diff --git a/.gitignore b/.gitignore
index 1d39c43..76a6b93 100644
--- a/.gitignore
+++ b/.gitignore
@@ -46,14 +46,18 @@ ENV/
 # ATM runtime artefacts
 logs/*.jsonl
 logs/dead_letter.jsonl
+logs/detections/
+logs/fires
+logs/pause.flag
 samples/*.png
 samples/*.jpg
 samples/labels.json
 trades.jsonl
 
-# configs: keep template + current marker, not generated calibration
+# configs: keep template only; ignore generated calibration and runtime state
 configs/*.toml
 !configs/example.toml
+configs/current.txt
 
 # Claude scheduler state
 .claude/
diff --git a/CLAUDE.md b/CLAUDE.md
new file mode 100644
index 0000000..ae8a8e3
--- /dev/null
+++ b/CLAUDE.md
@@ -0,0 +1,52 @@
+# ATM — Automated Trading Monitor
+
+Personal Faza-1 tool for the M2D strategy. Python 3.11+.
+
+## Quick Reference
+
+```bash
+pip install -e ".[windows]"    # Windows: live capture
+pip install -e ".[dev]"        # Linux/macOS: dev + tests (WSL: create venv first)
+atm calibrate                  # Tk wizard
+atm debug --delay 5            # one-shot capture + detect
+atm validate-calibration samples/calibration_labels.json   # offline color gate
+atm run --start-at 16:30 --stop-at 23:00                   # live session
+atm run --tz America/New_York --oh-start 09:30 --oh-stop 16:00  # NYSE window override
+atm dryrun samples             # corpus gate
+pytest -q                      # 184 tests
+```
+
+## Telegram commands (live)
+
+`/ss` `/status` `/pause` `/resume` `/resume force` `/3` (interval min) `/stop`
+
+- `/resume` clears only user pause; Canary drift requires `/resume force`.
+- Drift-pause now emits a single Telegram alert (was silent pre-refactor — root cause of the 2026-04-17 hang).
+
+## Operating-hours config
+
+`[options.operating_hours]` in TOML: `enabled`, `timezone` (NYSE local, e.g. `America/New_York`), `weekdays`, `start_hhmm`, `stop_hhmm`. Timezone validated at load; `_tz_cache` reused per tick. Boundary crossings log `market_open` / `market_closed` and notify once. Startup in-window is silent.
+
+## Phase-skip backstop
+
+`[options.alerts] fire_on_phase_skip = true` (default) — ARMED→light_*  direct (dark_* missed) still emits a `⚠️ PHASE SKIP` alert using FSM lockout to suppress spam.
+
+## Skill routing
+
+When the user's request matches an available skill, ALWAYS invoke it using the Skill
+tool as your FIRST action. Do NOT answer directly, do NOT use other tools first.
+The skill has specialized workflows that produce better results than ad-hoc answers.
+
+Key routing rules:
+- Product ideas, "is this worth building", brainstorming → invoke office-hours
+- Bugs, errors, "why is this broken", 500 errors → invoke investigate
+- Ship, deploy, push, create PR → invoke ship
+- QA, test the site, find bugs → invoke qa
+- Code review, check my diff → invoke review
+- Update docs after shipping → invoke document-release
+- Weekly retro → invoke retro
+- Design system, brand → invoke design-consultation
+- Visual audit, design polish → invoke design-review
+- Architecture review → invoke plan-eng-review
+- Save progress, checkpoint, resume → invoke checkpoint
+- Code quality, health check → invoke health
diff --git a/README.md b/README.md
index 2190508..2408c12 100644
--- a/README.md
+++ b/README.md
@@ -1,98 +1,117 @@
-# ATM — Automated Trading Monitor
+# ATM — Monitor Automat de Trading
 
-Personal Faza-1 tool for the **M2D strategy**. Watches the M2D MAPS colored-dot strip on a TradeStation chart, runs a phased state machine (ARMED→PRIMED→FIRE), pushes Discord + Telegram alerts with an annotated screenshot on BUY/SELL. You execute the trade manually in TradeLocker.
+Tool personal pentru strategia **M2D**. Urmărește banda de puncte colorate M2D MAPS de pe un chart TradeStation, rulează o mașină de stări pe faze (ARMED → PRIMED → FIRE) și trimite alerte pe Discord + Telegram cu screenshot adnotat la fiecare semnal BUY/SELL. **Execuția trade-ului o faci tu manual în TradeLocker.**
 
-No auto-execution. Faza 2 (auto-execute) is blocked on prop-firm TOS audit — see `docs/phase2-prop-firm-audit.md`.
+Fără execuție automată. Faza 2 (auto-execute) e blocată de auditul TOS prop-firm — vezi `docs/phase2-prop-firm-audit.md`.
 
 ---
 
-## Project layout
+## Cum e organizat proiectul
 
 ```
 atm/
-├── configs/              # calibration outputs + current.txt marker
+├── configs/              # calibrări + current.txt (marcaj care config e activ)
 ├── logs/
-│   ├── YYYY-MM-DD.jsonl  # per-cycle audit log, rotates at local midnight
-│   ├── dead_letter.jsonl # alerts that failed after retries
-│   ├── fires/            # annotated screenshots, one per BUY/SELL trigger
-│   └── calibrate_capture_*.png / debug_*.png  # gitignored debug artifacts
-├── samples/              # full frames saved automatically on each colour change
-├── src/atm/              # package
-│   ├── config.py         # frozen dataclass + TOML loader
-│   ├── vision.py         # ROI crop, phash, pixel↔price, Hough, connected-components
-│   ├── state_machine.py  # 5-state phased FSM, per-direction lockout
-│   ├── detector.py       # capture → crop → find rightmost dot → classify → debounce
-│   ├── canary.py         # layout phash drift watchdog with pause-file gating
-│   ├── levels.py         # Phase-B SL/TP line extraction
-│   ├── notifier/         # FanoutNotifier + Discord webhook + Telegram bot
-│   ├── audit.py          # line-buffered JSONL, daily rotation
-│   ├── calibrate.py      # Tk wizard (region-select + click-sample)
-│   ├── labeler.py        # Tk UI → labels.json
-│   ├── dryrun.py         # replay corpus, precision/recall gate
-│   ├── journal.py        # trade entries
-│   ├── report.py         # weekly R-multiple PnL
-│   └── main.py           # unified CLI
-├── tests/                # 105 pytest cases
-└── TODOS.md              # P1/P2/P3 backlog, Faza 2 items
+│   ├── YYYY-MM-DD.jsonl  # audit zilnic, se rotește la miezul nopții local
+│   ├── dead_letter.jsonl # alerte care au eșuat după retries
+│   ├── fires/            # screenshot-uri adnotate, unul per trigger BUY/SELL
+│   └── calibrate_capture_*.png / debug_*.png  # artefacte debug (gitignored)
+├── samples/              # frame complet salvat automat la fiecare schimbare de culoare
+├── src/atm/              # pachetul Python
+│   ├── config.py         # dataclass + loader TOML
+│   ├── vision.py         # crop ROI, phash, pixel↔preț, Hough, componente conectate
+│   ├── state_machine.py  # FSM 5 stări + lockout per direcție
+│   ├── detector.py       # capture → crop → găsește dot-ul rightmost → clasifică → debounce
+│   ├── canary.py         # watchdog layout via phash drift + flag de pauză
+│   ├── levels.py         # extracție SL/TP pe Faza-B
+│   ├── notifier/         # FanoutNotifier + webhook Discord + bot Telegram
+│   ├── audit.py          # JSONL line-buffered, rotație zilnică
+│   ├── calibrate.py      # wizard Tk (selectează regiune + click pe culori)
+│   ├── labeler.py        # UI Tk → labels.json
+│   ├── dryrun.py         # replay pe corpus, gate precision/recall
+│   ├── validate.py       # gate offline de clasificare a culorilor
+│   ├── journal.py        # înregistrări trade-uri
+│   ├── report.py         # raport săptămânal PnL în R
+│   └── main.py           # CLI unificat
+├── tests/                # 184 teste pytest
+└── TODOS.md              # backlog P1/P2/P3
 ```
 
 ---
 
-## Install
+## Instalare
 
-Python 3.11+ required. Clone, then:
+Python 3.11+.
 
 ```bash
-pip install -e ".[windows]"    # Windows: live capture + window focus
-pip install -e .               # Linux / macOS: dev / dryrun only (no live)
+pip install -e ".[windows]"    # Windows: capture live + focus fereastră
+pip install -e ".[dev]"        # Linux/macOS/WSL: doar dev + teste (fără capture)
 atm --help
 ```
 
-`[windows]` pulls `mss`, `pygetwindow`, `pywin32`.
+**WSL/Linux:** recomandat să folosești un virtualenv local:
+```bash
+python3 -m venv .venv
+source .venv/bin/activate
+pip install -e ".[dev]"
+```
+
+`[windows]` aduce `mss`, `pygetwindow`, `pywin32` (nu le pune pe WSL).
 
 ---
 
-## Calibration
+## Calibrare
 
-One-time per chart layout. Run on the machine that will do live capture.
+Se face o singură dată per layout de chart. Trebuie să ruleze pe mașina pe care face capture live (Windows, fizic — nu RDP/virtual).
 
 ```powershell
-atm calibrate           # 3s default countdown; use --delay 10 if you want more time
+atm calibrate           # countdown 3s default; pune --delay 10 dacă vrei mai mult timp
 ```
 
 Flow:
-1. Dialog: substring of the chart window title (e.g. `TradeStation` or `DIA`). Stored in config for later auto-focus.
-2. **"Ready?" message** → click OK → 3s countdown in terminal. Alt-tab TradeStation to the foreground and minimize anything covering it.
-3. Full-desktop screenshot is captured and shown in a scaled Tk window.
-4. **Drag a rectangle** over the chart (include the M2D MAPS strip). Enter = confirm. Esc = cancel.
-5. Step-by-step clicks on the selected region:
-   - M2D MAPS strip: top-left + bottom-right corners
-   - One click on each of: turquoise, yellow, dark_green, dark_red, light_green, light_red, gray dot + chart background (8 total — "Skip" if a colour isn't currently visible)
-   - Chart area: top-left + bottom-right (for Phase-B line detection)
-   - Two known price levels on the y-axis (pixel y → enter price)
-   - Canary region: top-left + bottom-right on a stable UI element (axis label, title bar)
-6. **Save** → writes `configs/YYYY-MM-DD-HHMM.toml` + marker `configs/current.txt`. Pulls Discord/Telegram creds from env (`ATM_DISCORD_URL`, `ATM_TG_TOKEN`, `ATM_TG_CHAT`) if set; otherwise `REPLACE_ME` placeholders — edit the TOML manually.
+1. Dialog: substring din titlul ferestrei chart-ului (ex. `TradeStation` sau `DIA`). Se salvează în config pentru auto-focus ulterior.
+2. **Mesaj "Ready?"** → click OK → countdown 3s în terminal. Alt-tab pe TradeStation, minimizează tot ce-l acoperă.
+3. Se face screenshot full-desktop, apare o fereastră Tk scalată.
+4. **Trage un dreptunghi** peste chart (include și banda M2D MAPS). Enter = confirmă. Esc = anulează.
+5. Click pas cu pas pe regiunea selectată:
+   - M2D MAPS strip: colț stânga-sus + colț dreapta-jos
+   - Un click pe fiecare culoare: turquoise, yellow, dark_green, dark_red, light_green, light_red, gray + background (8 total — "Skip" dacă o culoare nu-i vizibilă acum)
+   - Chart: colț stânga-sus + colț dreapta-jos (pentru detecția de linii în Faza-B)
+   - Două prețuri cunoscute pe axa Y (pixel y → introduci prețul)
+   - Canary: colț stânga-sus + colț dreapta-jos pe un element UI **stabil** (etichetă axă, bară titlu)
+6. **Save** → scrie `configs/YYYY-MM-DD-HHMM.toml` + marcaj `configs/current.txt`. Preia credențialele Discord/Telegram din env (`ATM_DISCORD_URL`, `ATM_TG_TOKEN`, `ATM_TG_CHAT`) dacă sunt setate; altfel pune `REPLACE_ME` — editezi TOML-ul manual.
 
-What gets written:
-- `chart_window_region = {x, y, w, h}` — virtual-desktop absolute rectangle. Runtime capture crops the same box, so the window must stay in that position.
-- `dot_roi`, `chart_roi`, `canary.roi` — coords relative to the selected region.
-- Per-colour RGB (sampled via saturation-snap within 15px of the click, mean of 5x5 around the snapped centre).
-- `y_axis` linear-interp pair.
-- `canary.baseline_phash` of the canary ROI.
+### ⚠️ Reguli critice la calibrare (evită incidentul 2026-04-17)
 
-Sampling tips:
-- Click colours that are **actually present** in the current dot history. If a colour isn't visible, skip it — `atm dryrun` will tell you if the skipped value doesn't match real dots.
-- Default tolerance is 60 for dot colours, 25 for background. Tighten via TOML after dryrun if misclassifications creep in.
+**1. Click EXCLUSIV pe dot-ul din DREAPTA al strip-ului.**
+Banda M2D MAPS e istoric: dot-ul din dreapta = activ/curent, restul sunt mai vechi. TradeStation desenează dot-ul activ mai strălucitor decât cele vechi. Detector-ul live citește MEREU dot-ul din dreapta. Dacă dai click pe unul din stânga, culoarea calibrată e mai întunecată decât realitatea → clasificare greșită live (dark_red poate ajunge citit ca light_red, de exemplu).
+
+**2. Canary pe un pixel STATIC.**
+NU pune regiunea canary peste: volume bar, preț curent, ceas/timestamp. Orice se schimbă natural în acea zonă declanșează drift-pause silent → bot-ul se oprește din detecție fără alertă vizibilă (asta s-a întâmplat la 22:25 pe 17.04, drift=129). Alege: o etichetă de axă, un titlu de panel, un colț de bordură.
+
+**3. Calibrează în mijlocul unei sesiuni active**, nu dimineața înainte de deschidere. Dot-urile sunt clar vizibile și reflectă exact aceleași setări de rendering ca la live.
+
+### Ce scrie în TOML
+
+- `chart_window_region = {x, y, w, h}` — dreptunghi absolut virtual-desktop. Capture-ul la runtime crop-ează exact aceeași cutie, deci fereastra **nu trebuie mutată** după calibrare.
+- `dot_roi`, `chart_roi`, `canary.roi` — coordonate relative la regiunea selectată.
+- RGB per culoare (eșantionat cu saturation-snap într-o rază de 15px de click, media unui box 5x5 în jurul pixelului snapped).
+- `y_axis` — pereche de interpolare liniară.
+- `canary.baseline_phash` al ROI-ului canary.
+
+Tips de sampling:
+- Click pe culori **chiar vizibile acum** în istoricul dot-urilor. Dacă o culoare nu-i vizibilă, skip — `atm dryrun` îți zice dacă valoarea ratată nu se potrivește cu dot-uri reale.
+- Tolerance default: 60 pentru dot-uri, 25 pentru background. Strângi în TOML după dryrun dacă apar misclasificări.
 
 ---
 
-## Smoke-test after calibration
+## Smoke-test după calibrare
 
 ```powershell
 atm debug --delay 5
 ```
 
-Captures one frame. Saves `logs/debug_full_<ts>.png`, `logs/debug_dot_roi_<ts>.png`, `logs/debug_annotated_<ts>.png`. Prints:
+Ia un frame. Salvează `logs/debug_full_<ts>.png`, `logs/debug_dot_roi_<ts>.png`, `logs/debug_annotated_<ts>.png`. Tipărește:
 
 ```
 window_found: True
@@ -102,116 +121,277 @@ classified:   gray  distance=24  confidence=0.79
 accepted:     True  color=gray
 ```
 
-Open the annotated PNG: yellow rectangle = `dot_roi`, red circle = detected dot. The circle should land on the ACTUAL rightmost colored dot in the M2D MAPS strip. If not:
-- Circle mid-strip → wrong window under the capture region (bring TradeStation to front).
-- Circle on a non-dot UI element → `dot_roi` boundaries capture too much; recalibrate narrower.
-- `color=None` + `UNKNOWN` → tolerances too tight OR sampled RGBs don't match real dots; recalibrate clicking on actual dots.
+Deschizi PNG-ul adnotat: dreptunghi galben = `dot_roi`, cerc roșu = dot detectat. Cercul trebuie să pice pe **dot-ul colorat cel mai din dreapta** din banda M2D MAPS. Dacă nu:
+- Cerc la mijloc de strip → alt window e sub regiunea de capture (adu TradeStation în față).
+- Cerc pe element UI non-dot → `dot_roi` prea larg; recalibrează mai îngust.
+- `color=None` + `UNKNOWN` → tolerances prea strânse SAU RGB-urile eșantionate nu se potrivesc cu dot-urile reale; recalibrează cu click pe dot-uri reale.
 
 ---
 
-## Live run
+## Validare offline a calibrării
+
+Verifici dacă calibrarea actuală clasifică corect un set de frame-uri etichetate manual, **fără să aștepți sesiunea live**. Esențial după orice recalibrare.
+
+```bash
+atm validate-calibration samples/calibration_labels.json
+```
+
+Format input (`samples/calibration_labels.json`):
+```json
+[
+  {"path": "logs/fires/20260417_201500_arm_sell.png", "expected": "yellow", "note": "primul arm"},
+  {"path": "logs/fires/20260417_205302_ss.png",       "expected": "dark_red"},
+  {"path": "logs/fires/20260417_210441_ss.png",       "expected": "light_red"}
+]
+```
+
+Output: per fiecare frame PASS/FAIL + culoarea detectată + top 3 candidați după distanță RGB + sugestii de pixel pentru misclasificări.
+
+Exit code:
+- `0` — 100% PASS (poți porni live în siguranță)
+- `1` — cel puțin un FAIL
+- `2` — input invalid/lipsă
+
+### Două corpus-uri, două scopuri
+
+| Corpus | Unde se salvează | Cum se populează | Folosit de |
+|---|---|---|---|
+| `samples/` | frame complet la fiecare **schimbare de culoare** detectată | automat de `atm run` | `atm label` + `atm dryrun` |
+| `logs/fires/` | screenshot adnotat la fiecare alertă BUY/SELL, `/ss` manual, **interval automat `/3`** | manual sau scheduler | `atm validate-calibration` |
+
+**Flow A — calibrare fină cu screenshots automate (`/3`)**
+
+Util când vrei să acumulezi repede frame-uri din culori reale, fără să aștepți schimbări de culoare.
+
+1. **În sesiunea live**, trimite `/3` în Telegram → bot-ul face screenshot automat la 3 minute și îl salvează în `logs/fires/*_ss.png`. Oprești cu `/stop`.
+2. **După sesiune**, adaugi intrări în `samples/calibration_labels.json` pentru fiecare screenshot relevant, cu culoarea pe care ai văzut-o TU pe chart:
+   ```json
+   {"path": "logs/fires/20260420_151234_ss.png", "expected": "dark_green", "note": "văzut live, ratat de bot"}
+   ```
+3. **Rulează validarea:**
+   ```bash
+   atm validate-calibration samples/calibration_labels.json
+   ```
+4. **Interpretează rezultatul:**
+   - **Toate PASS** → calibrarea ține, continui live fără modificări.
+   - **Măcar un FAIL** → output-ul îți arată pixelul real (ex. `RGB(128, 0, 0)`), centrul curent din TOML (ex. `dark_red RGB(83, 0, 0)`) și distanța. Două opțiuni:
+     - **Fix tactic rapid:** editezi TOML-ul direct, muți centrul culorii aproape de pixelul observat. Rulezi iar `validate-calibration`. Te oprești când e PASS.
+     - **Fix complet:** la următoarea sesiune live completă, rulezi `atm calibrate` de la zero pe Windows, cu **disciplina cele 3 reguli critice de mai sus** (rightmost dot, pixel static pentru canary, în timpul unei sesiuni active).
+5. **Acumulezi mai multe samples în timp.** Obiectiv: 2-3 intrări per culoare în `calibration_labels.json`. Cu cât fișierul are mai multe etichete, cu atât calibrarea următoare e validată mai solid.
+
+**Flow B — gate de precizie pe corpus de schimbări de culoare**
+
+`atm run` salvează automat în `samples/` un frame complet la fiecare schimbare de culoare detectată. După sesiune:
 
 ```powershell
-# Today's session 16:30–23:00 Romania local
+atm label samples     # UI Tk — etichetezi fiecare frame cu culoarea reală văzută pe chart
+atm dryrun samples    # replay prin detector + FSM; exit 0 dacă precision=100%, recall≥95%
+```
+
+Dacă gate-ul pică, ajustezi `tolerance` per culoare în TOML sau corectezi eșantioanele nepotrivite, apoi rulezi iar `atm dryrun` până trece.
+
+### Workflow de corectare iterativă (când apare o alertă greșită live)
+
+Scenariu: ai rulat o sesiune live, ai văzut pe chart o culoare pe care bot-ul n-a detectat-o (sau a detectat greșit).
+
+1. **În timpul sesiunii** — două opțiuni pentru a captura dovezi:
+   - `/ss` în Telegram → un screenshot instant în `logs/fires/`
+   - `/3` în Telegram → screenshots automate la 3 min în `logs/fires/` (util dacă nu ești la monitor continuu); oprești cu `/stop`
+2. **După sesiune**, adaugi intrările relevante în `samples/calibration_labels.json` cu culoarea corectă și rulezi `atm validate-calibration` (Flow A de mai sus).
+3. Dacă apar FAIL-uri, aplici fix tactic în TOML sau recalibrezi complet.
+
+### Exemplu real — incidentul 2026-04-17
+
+La 20:53 s-a afișat un dark_red pe chart dar bot-ul l-a citit ca light_red (alertă ratată). Root cause: calibrarea anterioară (`2026-04-16-0703.toml`) a fost făcută dând click pe dot-uri istorice (mai întunecate), nu pe dot-ul activ din dreapta.
+
+Fix aplicat în `2026-04-18-1220.toml`, pe bază de evidență live:
+
+| Culoare | Centru vechi | Pixel live observat | Centru nou |
+|---|---|---|---|
+| dark_red | (83, 0, 0) | (128, 0, 0) | **(128, 0, 0)** |
+| light_red | (153, 0, 0) | (171, 0, 0) | **(171, 0, 0)** |
+| dark_green | (0, 77, 0) | — | **(0, 122, 0)** (ajustat proporțional: +45 pe G) |
+| light_green | (0, 153, 0) | — | **(0, 171, 0)** (ajustat proporțional: +18 pe G) |
+
+yellow, turquoise, gray, background — lăsate neschimbate (nu am dovezi live care să justifice ajustarea).
+
+După fix: `atm validate-calibration` → 3/3 PASS, confidence 1.00 pe ambele roșuri.
+
+**Rollback** dacă ceva merge prost:
+```bash
+echo "2026-04-16-0703.toml" > configs/current.txt
+```
+
+---
+
+## Sesiunea live
+
+```powershell
+# Sesiunea de azi 16:30–23:00 România local
 atm run --start-at 16:30 --stop-at 23:00
 
-# Indefinite
+# Fără limită
 atm run
 
-# Fixed duration (hours)
+# Durată fixă (ore)
 atm run --duration 2
 
-# Linux / headless smoke (reads samples/*.png in a loop)
+# Linux/WSL smoke (rulează pe fișiere din samples/)
 atm run --capture-stub --duration 0.05
 ```
 
-Startup sequence:
-1. Wall-clock wait until `--start-at` (if set).
-2. `pygetwindow.activate()` on the first window matching `cfg.window_title` — brings TradeStation to the foreground automatically (restores if minimised).
-3. 5s countdown (`--startup-delay`).
-4. Capture first frame + canary check. Status (`drift=X/Y` or `capture_failed`) is included in the startup ping.
-5. **"ATM started" ping** on Discord + Telegram.
-6. Main loop: every `loop_interval_s` (default 5s) — capture → canary → detect → state machine → maybe notify → maybe Phase-B.
-7. At `--stop-at` (or `--duration`): **"ATM stopped" ping**, then exit.
+Startup:
+1. Așteptare wall-clock până la `--start-at` (dacă e setat).
+2. `pygetwindow.activate()` pe prima fereastră care conține `cfg.window_title` — aduce TradeStation în față (restaurează dacă-i minimizată).
+3. Countdown 5s (`--startup-delay`).
+4. Primul frame + check canary. Status (`drift=X/Y` sau `capture_failed`) e inclus în ping-ul de start.
+5. **Ping "ATM started"** pe Discord + Telegram.
+6. Loop principal: la fiecare `loop_interval_s` (default 5s) — capture → canary → detect → FSM → poate notifică → poate Faza-B.
+7. La `--stop-at` (sau `--duration`): **ping "ATM stopped"**, apoi exit.
 
-Per-cycle behaviour:
-- Canary drift → auto-pause (logs `paused`, skips detection). Clear by running `atm run` again with the pause-file removed.
-- Detector reports UNKNOWN → stays in current state (logged as `noise`).
-- Colour change → full frame saved to `samples/YYYYMMDD_HHMMSS_<color>.png` (for corpus).
-- FIRE (BUY/SELL, not locked) → annotated PNG saved to `logs/fires/`, attached to the alert, `LevelsExtractor` armed.
-- Phase-B complete → "Levels SL=… TP1=… TP2=…" push.
-- Heartbeat every `heartbeat_min` minutes.
+Comportament per ciclu:
+- Drift canary → auto-pause + **alertă Telegram single-shot** (`⚠️ Canary drift=N — monitorizare pauzată`). Anulezi cu `/resume force` în Telegram, sau repornești cu flag-ul de pauză șters.
+- Detector raportează UNKNOWN → rămâne în starea curentă (loghează `noise`).
+- Schimbare de culoare → frame complet salvat în `samples/YYYYMMDD_HHMMSS_<color>.png` (pentru corpus).
+- FIRE (BUY/SELL, nu locked) → PNG adnotat salvat în `logs/fires/`, atașat la alertă, `LevelsExtractor` armed.
+- **Phase-skip backstop** (`fire_on_phase_skip=true` default) → ARMED → light_red/light_green direct (dark_* ratat) emite totuși alertă `⚠️ PHASE SKIP` cu screenshot. Lockout-ul FSM previne spam.
+- Faza-B completă → push "Levels SL=… TP1=… TP2=…".
+- Heartbeat la fiecare `heartbeat_min` minute.
 
-Keep PowerShell minimized during the session so it doesn't cover TradeStation.
+Ține PowerShell minimizat în timpul sesiunii ca să nu acopere TradeStation.
+
+### Fereastra orelor de trading
+
+Configurezi din TOML (sursă adevăr: NYSE local, timezone-aware — DST-ul e gestionat automat):
+
+```toml
+[options.operating_hours]
+enabled     = true
+timezone    = "America/New_York"     # validat fail-fast la load
+weekdays    = ["MON", "TUE", "WED", "THU", "FRI"]
+start_hhmm  = "09:30"                # deschidere NYSE
+stop_hhmm   = "16:00"                # închidere NYSE
+```
+
+Tick-urile din afara ferestrei sunt skipped (logged doar la tranziție). La traversarea boundary-ului bot-ul emite `market_open` / `market_closed` în Telegram — o singură dată per tranziție. **Pornirea în-fereastră nu emite alertă spurioasă.**
+
+Override din CLI (bat TOML-ul):
+
+```
+atm run --tz America/New_York --weekdays MON,TUE,WED,THU,FRI --oh-start 09:30 --oh-stop 16:00
+```
+
+> `--oh-start / --oh-stop` sunt **diferite** de `--start-at / --stop-at`.
+> `--start-at / --stop-at` = wall-clock session bounds (când pornește procesul și când se oprește).
+> `--oh-start / --oh-stop` = fereastra NYSE în care detecția rulează efectiv în interiorul sesiunii.
+> Se combină.
+
+### Comenzi Telegram
+
+Trimiți în chat-ul bot-ului:
+
+| Comandă | Efect |
+|---|---|
+| `/ss` sau `/screenshot` | Screenshot acum |
+| `/status` | Stare FSM + motiv pauză + fereastră open/closed |
+| `/pause` | Suspendă detecția (heartbeat-urile continuă) |
+| `/resume` | Elimină DOAR pauza user. Dacă Canary e drift-paused, **rămâne paused** — folosește `/resume force` |
+| `/resume force` | Elimină și drift-pause-ul canary (după recalibrare) |
+| `/3` sau `/interval 3` | Interval auto-screenshot = 3 min |
+| `/stop` | Oprește scheduler-ul de screenshot |
+
+Doar `allowed_chat_ids` sunt acceptate. După 3 `401` consecutive, poller-ul intră în mod degradat.
 
 ---
 
-## After the session
+## După sesiune
 
 ```powershell
-atm label samples     # Tk UI — label each saved frame with true dot colour
-atm dryrun samples    # replay through detector + FSM; exits 0 if precision=100%, recall>=95%
+atm label samples     # UI Tk — etichetezi fiecare frame salvat cu culoarea reală
+atm dryrun samples    # replay prin detector + FSM; exit 0 dacă precision=100%, recall≥95%
 ```
 
-If the gate fails, tune per-colour `tolerance` in `configs/<current>.toml`, or recalibrate colour samples that didn't match. Re-run `atm dryrun` until it passes. Only then do you trust live signals.
+Dacă gate-ul pică, ajustezi `tolerance` per culoare în `configs/<current>.toml`, sau recalibrezi eșantioanele care n-au potrivit. Rulezi iar `atm dryrun` până trece. **Numai atunci ai încredere în semnalele live.**
 
-Trade record-keeping:
+Pentru calibrare fină a clasificării de culori (Flow A cu `/3`), vezi secțiunea **Validare offline a calibrării** de mai sus.
+
+Evidență trade-uri:
 
 ```powershell
-atm journal                      # interactive entry after a real trade
-atm report --week 2026-16        # weekly win rate + R PnL + slippage
+atm journal                      # înregistrare interactivă după un trade real
+atm report --week 2026-16        # win rate săptămânal + PnL în R + slippage
 ```
 
 ---
 
-## DPI / multi-monitor notes
+## Note DPI / multi-monitor
 
-- Calibration region is virtual-desktop-absolute; runtime capture uses the same rectangle. **Don't move the TradeStation window** after calibrating. Canary will catch drift and pause automatically.
-- Changing DPI scaling or moving to a different monitor with different DPI → recalibrate.
-- RDP / virtual desktops: `mss` can return black frames over RDP. Run locally on the same physical machine as TradeStation.
+- Regiunea din calibrare e absolută virtual-desktop; runtime capture folosește același dreptunghi. **Nu muta fereastra TradeStation** după calibrare. Canary prinde drift-ul și pauzează automat.
+- Schimbi DPI scaling sau muți pe un alt monitor cu DPI diferit → recalibrezi.
+- RDP / desktop virtual: `mss` poate returna frame-uri negre peste RDP. Rulează local pe aceeași mașină fizică pe care e TradeStation.
 
 ---
 
 ## Troubleshooting
 
-| Symptom | Likely cause | Fix |
+| Simptom | Cauză probabilă | Fix |
 |---|---|---|
-| `capture_failed` in startup ping | `chart_window_region` references coords off-screen (different monitor layout) | Recalibrate. |
-| Startup canary `drift=X/8` with X >> 8 | Wrong window is in the capture region | Make sure TradeStation is the window at `cfg.chart_window_region`. Relaunch. |
-| `WARN: no window contains 'xxx'` at startup | `cfg.window_title` substring matches nothing | Edit `window_title` in TOML to a substring that's unique to TradeStation. |
-| No alerts even after trigger ought to fire | Check `logs/YYYY-MM-DD.jsonl` for `event=tick` entries — are colours accepted? Is `trigger` ever set? | If always UNKNOWN → tolerances too tight. If `trigger` but `locked=true` → lockout from prior fire, normal. |
-| Discord OK, Telegram silent (or vice versa) | `logs/dead_letter.jsonl` contains failed alerts with error | Fix credentials in TOML, restart. |
-| Heartbeat shows `telegram: failed > 0` | Telegram returned `ok:false` (bot blocked, invalid chat_id, parse error) | Check `logs/dead_letter.jsonl` for the `error_str` / `description` field. Common: bot never started by user in Telegram, or wrong `chat_id` flavor (channel vs group vs DM). |
-| Debug circle on mid-strip instead of right edge | Anti-aliasing bridges dots in the mask | Already fixed via erosion+connected-components — ensure `git pull` is current. |
-| Wizard window is tiny / image not visible | Tk geometry default on Windows | Already fixed — `git pull`. Image is scaled to fit screen. |
+| `capture_failed` în ping-ul de start | `chart_window_region` referă coords off-screen (alt layout monitor) | Recalibrează. |
+| Canary la startup arată `drift=X/8` cu X ≫ 8 | Alt window e în regiunea de capture | TradeStation trebuie să fie ferestra la `cfg.chart_window_region`. Relansează. |
+| `WARN: no window contains 'xxx'` la start | `cfg.window_title` nu prinde nimic | Editează `window_title` în TOML cu un substring unic pentru TradeStation. |
+| Nu vin alerte deși ar trebui | Verifică `logs/YYYY-MM-DD.jsonl` — `event=frame` au culoare acceptată? `trigger` setat? | Dacă mereu UNKNOWN → tolerances prea strânse SAU RGB-urile calibrate nu se potrivesc. Rulează `atm validate-calibration`. Dacă `trigger` dar `locked=true` → lockout de la fire anterior, normal. |
+| Alertă pe culoare greșită (ex. dark_red → light_red) | Calibrarea a luat dot istoric, nu activ | Rulează `atm validate-calibration`. Corectezi tactic în TOML sau recalibrezi cu regula rightmost dot. |
+| Discord OK, Telegram tace (sau invers) | `logs/dead_letter.jsonl` are alertele eșuate + eroarea | Fixezi credențiale în TOML, restart. |
+| Heartbeat arată `telegram: failed > 0` | Telegram a răspuns `ok:false` | Check `logs/dead_letter.jsonl` pentru `error_str` / `description`. Comun: bot-ul nu-a fost pornit de user în Telegram, sau `chat_id` greșit (channel vs group vs DM). |
+| Bot-ul "moare" după N ore, heartbeat merge dar comenzile nu răspund | Era bug-ul de hang din 2026-04-17 — drain coadă de comenzi sărit când Canary paused | Fixat în `c5024ce`. Update git pull. |
 
 ---
 
-## Windows Task Scheduler (production)
+## Windows Task Scheduler (producție)
 
-For hands-off daily runs surviving reboots:
+Pentru rulare automată zilnică care supraviețuiește reboot-urilor:
 
-1. Task Scheduler → Create Task → name `ATM M2D Monitor`
+1. Task Scheduler → Create Task → nume `ATM M2D Monitor`
 2. **General**: "Run only when user is logged on", "Run with highest privileges"
 3. **Triggers**: New → Daily, Start `16:30`
 4. **Actions**: New → Program `C:\path\to\python.exe`, Arguments `-m atm run --stop-at 23:00`, Start in `D:\PROIECTE\atm`
-5. **Conditions**: uncheck "Start only if AC power" (if laptop)
+5. **Conditions**: debifează "Start only if AC power" (dacă e laptop)
 6. **Settings**: "If task runs longer than 7 hours → stop"
 
-Click-right → Run to test manually. Manual DST-change check twice a year (Mar / Oct first week).
+Click-right → Run, să testezi manual. Check DST schimbare de două ori pe an (prima săptămână din martie / octombrie).
 
 ---
 
-## Quick command reference
+## Referință rapidă comenzi
 
 ```
-atm calibrate [--screenshot PATH] [--delay SEC]      # Tk wizard
+atm calibrate [--screenshot PATH] [--delay SEC]      # wizard Tk
 atm debug [--delay SEC]                              # one-shot capture + detect
-atm label SAMPLES_DIR                                # Tk labeling
-atm dryrun SAMPLES_DIR                               # corpus gate
+atm label SAMPLES_DIR                                # etichetare Tk
+atm dryrun SAMPLES_DIR                               # gate pe corpus
+atm validate-calibration LABEL_FILE.json             # gate offline clasificare culori
 atm run [--duration H] [--start-at HH:MM] [--stop-at HH:MM] [--startup-delay SEC] [--capture-stub]
-atm journal [--file PATH]                            # interactive trade entry
-atm report [--week YYYY-WW] [--file PATH]            # weekly summary
+        [--tz TZNAME] [--weekdays MON,TUE,...] [--oh-start HH:MM] [--oh-stop HH:MM]
+atm journal [--file PATH]                            # înregistrare interactivă
+atm report [--week YYYY-WW] [--file PATH]            # raport săptămânal
 ```
 
-Exit code: `atm dryrun` exits 0 if gate passes, 1 otherwise. Other commands follow standard convention.
+Exit codes:
+- `atm dryrun` — 0 pass, 1 fail.
+- `atm validate-calibration` — 0 toate PASS, 1 orice FAIL, 2 input invalid.
+- Restul: standard.
+
+---
+
+## Evenimente audit
+
+Scrise în `logs/YYYY-MM-DD.jsonl`. Cele adăugate recent:
+
+| Event | Payload | Când |
+|---|---|---|
+| `canary_drift_paused` | `distance` | Primul tick cu drift după o stare curată; emite alertă Telegram |
+| `user_paused` | — | `/pause` primit |
+| `user_resumed` | `was_drift`, `was_user`, `force` | `/resume` sau `/resume force` |
+| `market_open` / `market_closed` | `reason` | Boundary fereastră operating-hours (o dată per tranziție; **nu** la startup) |
+| `phase_skip_fire` | `direction` | Alertă backstop când ARMED→light_* direct |
+| `command_error` | `action`, `error` | Excepție la dispatch (izolată de loop-ul de detecție) |
diff --git a/TODOS.md b/TODOS.md
index 120a8af..a142cf0 100644
--- a/TODOS.md
+++ b/TODOS.md
@@ -49,9 +49,23 @@ Read-only web view of today's audit JSONL + recent triggers. Useful for review a
 
 ---
 
+## P2-yaxis-recalib-detect — Y-axis recalibration detection
+
+Price overlay (from Telegram commands feature) uses `y_axis` linear interpolation to show current price on screenshots. When the user rescales the chart y-axis (common after overnight price gaps), the calibration becomes stale and prices shown are incorrect. Canary check detects layout drift but NOT y-axis range changes.
+
+- Possible approaches: OCR on y-axis labels (fragile), track price range consistency across sessions, or simple "calibration age" warning after N hours.
+- Start after price overlay is live and the false-price frequency is known.
+- Depends on: Telegram commands + price overlay feature being shipped.
+
 ## Quality debt
 
-- [ ] **Integration test for run_live loop**: currently mocked at module level. Add a short-duration in-memory loop test that threads real detector/state_machine/audit together (no network).
+- [x] **Integration test for run_live loop**: lifecycle async test added in `tests/test_main.py` (IDLE→ARMED→PRIMED auto-poll→FIRE auto-stop).
+- [x] **Detection-loop hang on canary pause** (2026-04-17 incident): `_drain_cmd_queue` now runs unconditionally; helpers extracted to module scope for testability (commit `c5024ce`).
+- [x] **Silent canary drift-pause**: single-shot Telegram alert on `not_paused → paused` (commit `9cf49ca`).
+- [x] **Phase-skip backstop**: `fire_on_phase_skip` (default on) emits alert when ARMED→light_* direct (commit `8b53b8d`).
+- [x] **Operating hours window**: NYSE-timezone-aware gate with `/pause` `/resume` `/resume force` control (commits `54f5575`, `2386577`).
+- [x] **Offline calibration gate**: `atm validate-calibration` replays labeled frames through detector (commit `8bae507`).
 - [ ] **Coverage report**: run `pytest --cov=atm --cov-report=term-missing`, aim for ≥ 85% per module.
 - [ ] **Typing strictness**: run `pyright src/` with strict mode, fix reported issues.
 - [ ] **Perf baseline**: profile one detection cycle on a representative frame; ensure < 100ms so 5s loop has ample headroom.
+- [ ] **Exchange calendar holidays**: operating_hours doesn't know about NYSE closures (MLK, Thanksgiving, Good Friday). User `/pause`s manually for now.
diff --git a/configs/example.toml b/configs/example.toml
index cdcb3c9..1602547 100644
--- a/configs/example.toml
+++ b/configs/example.toml
@@ -81,6 +81,24 @@ low_conf_run = 3
 phaseb_timeout_s = 600
 dead_letter_path = "logs/dead_letter.jsonl"
 
+# Alert-behavior toggles (not screenshot-attachment; see attach_screenshots below).
+# fire_on_phase_skip: emit a backstop "PHASE SKIP" alert when the FSM observes
+# ARMED → light_green/light_red directly (skipping the dark prime). Default on
+# because missing a fire is worse than a false-positive phase-skip alert.
+[options.alerts]
+fire_on_phase_skip = true
+
+# Operating hours — detection only runs on allowed weekdays + HH:MM window.
+# Timezone is the source of truth (NYSE local); the runtime converts tick
+# timestamps to this zone so DST rollovers stay aligned with the exchange.
+# Override from CLI with --tz / --weekdays / --oh-start / --oh-stop.
+[options.operating_hours]
+enabled = false
+timezone = "America/New_York"
+weekdays = ["MON", "TUE", "WED", "THU", "FRI"]
+start_hhmm = "09:30"
+stop_hhmm = "16:00"
+
 # Per-kind screenshot-attach toggles. All default to true on upgrade.
 # Accepts either a bare bool (legacy: attach_screenshots = true) or this table.
 [options.attach_screenshots]
diff --git a/docs/claude-master-design-20260415-atm-trading.md b/docs/claude-master-design-20260415-atm-trading.md
deleted file mode 100644
index 6906a43..0000000
--- a/docs/claude-master-design-20260415-atm-trading.md
+++ /dev/null
@@ -1,149 +0,0 @@
-# Design: ATM — Automated Trading Monitor (M2D Strategy)
-
-Generated by /office-hours on 2026-04-15
-Branch: master
-Repo: /workspace/atm (greenfield)
-Status: APPROVED
-Mode: Builder (personal live-trading tool, high-stakes)
-
-## Problem Statement
-
-User trades the M2D strategy on DIA (TradeStation chart with custom indicator) with execution on TradeLocker US30 CFD (prop firm account). Same strategy also applies to GLD → XAUUSD. Bridging signal source (TradeStation Windows app) with execution (TradeLocker web) currently requires user to watch both screens for 4 hours per evening. Goal: bot detects the trigger signal automatically and notifies user via Telegram/Discord with chart screenshot + SL/TP levels so user can execute the trade in TradeLocker.
-
-## Strategy M2D — Full Spec
-
-**Setup:** TradeStation, 3-minute chart, DIA (or GLD) symbol, custom indicator "M2D MAPS" that renders a horizontal strip of colored dots below the price panel. Dots are indexed by time, y-position is fixed.
-
-### BUY sequence (sequential in time, rightmost N dots):
-1. **Turquoise dot** — 15-minute buy trigger
-2. **Dark green dot** — 3-minute sell
-3. **Light green dot** — 3-minute buy → **TRIGGER**
-
-At trigger:
-- Execute BUY on TradeLocker, instrument US30 CFD
-- Stop Loss 0.6%
-- Volume 0.1 lots maximum
-- TP1, TP2, SL are drawn automatically as horizontal lines on the TradeStation chart after entry
-- User manual lifecycle: at TP1 close half, move SL to ~breakeven; at TP2 close remaining half
-
-### SELL sequence (mirror):
-1. **Yellow dot** — 15-minute sell (red 15min candle)
-2. **Dark red dot** — 3-minute buy
-3. **Light red dot** — 3-minute sell → **TRIGGER**
-
-Same size (0.1 lots), same SL %, same TP management.
-
-### Instrument mapping (intentional asymmetry):
-- DIA chart (TradeStation) ↔ US30 CFD (TradeLocker)
-- GLD chart (TradeStation) ↔ XAUUSD CFD (TradeLocker)
-
-### Trading window:
-- NY open first 2 hours + NY close last 2 hours
-- RO summer time: 16:30-18:30 and 21:00-23:00
-- Typical frequency: 1 trade per evening
-
-## Constraints
-
-- **Prop firm account on TradeLocker.** Faza 2 (auto-execution) requires reading prop TOS first — many prop firms prohibit automation or detect robotic timing patterns.
-- No API on TradeLocker. No signal export on TradeStation for compiled custom indicator.
-- Bot runs on the same Windows machine as TradeStation. Cross-machine (RDP/VNC) screenshot adds latency and fragility.
-
-## Premises (agreed)
-
-1. Screenshot + visual detection is the only viable bridge.
-2. Notification-first (Faza 1) is the right sequencing. Zero-click MVP removes all financial bug risk.
-3. M2D MAPS dot strip has stable y-position on fixed TradeStation layout → ROI color sampling is the right detection method.
-4. DIA→US30 price divergence is acceptable risk (user's judgment, has been trading this pairing live).
-5. Bot runs on the same Windows machine as TradeStation.
-
-## Recommended Approach — B: Structured Service with Dry-Run and Audit Log
-
-Python package on Windows, structured for clean extension to Faza 2.
-
-### Components:
-- **Detector core:** `mss` screenshot of TradeStation window (located by title via `pygetwindow`) → crop M2D MAPS ROI → scan rightmost N dot positions → classify each by closest-color match with tolerance → feed into state machine that tracks 3-dot sequences (turquoise→dark-green→light-green = BUY trigger; yellow→dark-red→light-red = SELL trigger).
-- **Level extractor:** after trigger, scan chart region for horizontal colored lines (SL/TP1/TP2). Convert pixel y to price via calibration of y-axis scale.
-- **Calibration tool (Tkinter):** interactive — user clicks on each dot color sample, captures RGB + tolerance, clicks on ROI corners, captures y-axis price references. Writes to `config.toml`.
-- **Dry-run mode:** runs detector against a folder of saved screenshots (recorded during normal operation). Shows what notification WOULD have been sent for each. Used to validate new color thresholds or strategy tweaks without live risk.
-- **Notifier abstraction:** interface with Discord webhook and Telegram bot implementations. Sends: annotated screenshot + decoded SL/TP1/TP2 prices + signal type (BUY/SELL) + timestamp.
-- **Audit log (JSONL):** every detection cycle — timestamp, detected dots, classification, decision, notification sent y/n. Replayable, debuggable.
-- **Scheduler:** Windows Task Scheduler entry, auto-start/stop at 16:30 / 18:30 / 21:00 / 23:00 local time (summer/winter offset aware).
-
-### Structure:
-```
-atm/
-├── pyproject.toml
-├── config.toml                  # populated by calibration tool
-├── src/atm/
-│   ├── detector.py              # screenshot + color classification + state machine
-│   ├── levels.py                # SL/TP1/TP2 pixel-to-price extraction
-│   ├── notifier/
-│   │   ├── __init__.py          # abstract Notifier
-│   │   ├── discord.py
-│   │   └── telegram.py
-│   ├── audit.py                 # JSONL logger
-│   ├── calibrate.py             # Tkinter UI
-│   ├── dryrun.py                # replay on saved screenshots
-│   └── main.py                  # orchestration + scheduler hooks
-├── samples/                     # saved screenshots for dry-run corpus
-└── logs/                        # JSONL audit
-```
-
-### Detection algorithm (core loop):
-1. Every 1 second during trading window:
-   - Locate TradeStation window
-   - If not foreground or minimized, log + skip
-   - Screenshot M2D MAPS ROI (fixed offsets from window bounds)
-   - For rightmost N=5 dot positions, sample center pixel, classify to nearest labeled color within tolerance
-   - Update rolling window of last 10 dots with their timestamps
-   - Evaluate state machine: did the last 3 classified dots (within a bounded time window) complete a BUY or SELL sequence?
-   - If trigger fired AND not already fired for this bar: extract SL/TP1/TP2 levels, send notification, log, mark fired.
-
-### Anti-duplicate logic:
-- Each trigger dot is keyed by (x-pixel position at capture, color). Once fired, stored in "recently fired" set with 10-minute TTL. Prevents re-fire if same dot persists across cycles.
-
-### Sanity guards:
-- If classification confidence (color distance) low for 3+ cycles in a row → push "bot lost sight" alert to user. Layout may have changed.
-- If TradeStation window not found for 60 seconds → push "bot cannot find chart" alert.
-
-## Open Questions (non-blocking)
-
-- Exact color tolerance values — determined during calibration session, not a design question.
-- GLD/XAUUSD: same M2D indicator on GLD chart? Assume yes, confirm during calibration.
-- Multi-symbol monitoring — single window switched manually, or two TradeStation windows side by side? Defer; v1 = single chart at a time, user switches manually.
-
-## Success Criteria (Faza 1)
-
-- Over 20 live trading sessions, bot detects ≥95% of signals user also spotted manually.
-- Zero false-positive notifications during the bot's first 5 sessions (tune tolerances aggressively).
-- Notification delivered within 3 seconds of trigger dot appearing.
-- Audit log lets user reproduce "why was no notification sent" for any missed signal.
-
-## Distribution Plan
-
-Personal tool, single user. No distribution channel needed — runs locally on user's Windows box. Git repo at `/workspace/atm`. `pyproject.toml` + `pip install -e .` for local dev. No CI/CD; user's own `scheduled task` starts/stops it.
-
-## Risk Flag — Faza 2 (deferred)
-
-Before extending to auto-execution in TradeLocker:
-1. Read prop firm TOS (search for "EA", "automation", "bot", "copy trading", "external signal"). If prohibited, **Faza 2 is off the table** — tool stays notification-only.
-2. If permitted, implement via Playwright browser automation against TradeLocker web UI.
-3. Add human-like click timing randomization (100-400ms jitter) to avoid robotic detection.
-4. Dry-run mode then becomes: "click coordinates resolved, action NOT sent" — user reviews the intended click before enabling live.
-
-## Next Steps (concrete)
-
-1. Init `/workspace/atm` as Python project. `pyproject.toml`, basic structure.
-2. Build calibration tool first. Without calibrated config, nothing works.
-3. Record 20-30 sample screenshots across several trading sessions (can start this today — doesn't need any code yet; just `mss` screenshot on a 5-second timer dumping to disk).
-4. Build detector + state machine. Validate against recorded screenshots in dry-run mode.
-5. Wire Discord webhook first (simpler than Telegram bot). Test end-to-end on live session.
-6. Add audit log.
-7. Schedule Windows task for trading hours.
-
-## What I noticed about how you think
-
-- You explicitly asked for dry-run before writing a line of code. "Să verific dacă vrea să apese corect, fără să apese efectiv." That's not a common instinct for someone building their own tool; it's the instinct of someone who has already had something break expensively.
-- You phased the project yourself — "faza 2 după ce mă conving că merge." That's the right ordering and you arrived at it unprompted.
-- When I challenged the API premise, you answered with specifics: the indicator is custom, the account doesn't support API. You knew the constraint, not guessed it.
-- You flagged the prop account almost casually at the end. A lot of builders would have skipped that detail. It turned out to be the most important constraint in the entire design.
diff --git a/docs/happy-swinging-mccarthy.md b/docs/happy-swinging-mccarthy.md
deleted file mode 100644
index 1d31520..0000000
--- a/docs/happy-swinging-mccarthy.md
+++ /dev/null
@@ -1,43 +0,0 @@
-# Plan: ATM Eng Review — Findings Applied
-
-## Context
-
-User ran `/plan-eng-review` on `partitioned-honking-unicorn.md` (ATM trading monitor, Faza 1). Eng review complete. All 4 decisions resolved, obvious fixes applied, plan file updated in place.
-
-## Where the changes live
-
-The reviewed plan (with all eng-review edits) is at:
-**`/home/claude/.claude/plans/partitioned-honking-unicorn.md`**
-
-Test plan artifact at:
-**`~/.gstack/projects/romfast-workspace/claude-master-eng-review-test-plan-20260415-212932.md`**
-
-## What changed in the reviewed plan
-
-### 4 decisions (AskUserQuestion)
-1. **Bar flicker** → debounce depth=1 (configurable); screenshot in alert = visual check.
-2. **Phase A entry price** → dropped; Phase A is direction + screenshot only; user puts manual 0.6% SL in TradeLocker; Phase B sends real levels from chart.
-3. **Notifier blocking** → fire-and-forget worker threads per backend, bounded queue, retry + dead-letter.
-4. **Alert SPoF** → Discord + Telegram parallel from day 1.
-
-### Obvious fixes (stated, applied)
-- Exhaustive state transition table (default-noise rule, SELL mirror explicit, phase-skip handling).
-- Python 3.11+ pin → drop `tomli`, use stdlib `tomllib`.
-- Windows symlink replaced by `configs/current.txt` marker file.
-- New `vision.py` shared module (ROI/hash/interp/Hough).
-- `@dataclass Config` with load-time validation.
-- DPI check added to calibrate + README note.
-
-### Test coverage
-Expanded from state-machine-only to: every module + 1 E2E replay harness. Acceptance gate unchanged (precision=100%, recall≥95% on labeled corpus).
-
-## Verification (post-implementation)
-
-Run the full verification checklist from `partitioned-honking-unicorn.md` (sections 1-9). Specifically:
-- `pytest tests/` — all new unit tests + E2E replay pass.
-- `atm dryrun ./samples` hits acceptance gate.
-- Live 2-session test: both Discord and Telegram fire; kill one mid-session and confirm the other still delivers + dead-letter file gets the failed alert.
-
-## Status
-
-**CEO + ENG CLEARED.** No further reviews required before implementation. Design + DX reviews properly skipped (no UI scope; personal single-user tool). Run `/ship` after implementation.
diff --git a/docs/image.png b/docs/image.png
deleted file mode 100644
index be291ad..0000000
Binary files a/docs/image.png and /dev/null differ
diff --git a/docs/partitioned-honking-unicorn.md b/docs/partitioned-honking-unicorn.md
deleted file mode 100644
index b2eca75..0000000
--- a/docs/partitioned-honking-unicorn.md
+++ /dev/null
@@ -1,258 +0,0 @@
-# Plan: ATM — Automated Trading Monitor (M2D, Faza 1) — ENG-REVIEWED
-
-**Source plan:** `/home/claude/.claude/plans/swirling-drifting-starfish.md`
-**CEO plan artifact:** `~/.gstack/projects/romfast-workspace/ceo-plans/2026-04-15-atm-trading.md`
-**Eng review mode:** FULL_REVIEW (4 decisions made, 0 unresolved)
-**Design doc:** `~/.gstack/projects/romfast-workspace/claude-master-design-20260415-atm-trading.md` (APPROVED)
-**Eng test plan:** `~/.gstack/projects/romfast-workspace/claude-master-eng-review-test-plan-20260415-212932.md`
-
----
-
-## Context
-
-User trades M2D strategy manually on DIA (TradeStation) with execution on TradeLocker US30 CFD (prop firm). Same strategy on GLD → XAUUSD. 4h/evening dual-screen monitoring. Faza 1 goal: bot auto-detects M2D trigger, sends Discord/Telegram notification with screenshot + SL/TP1/TP2 levels; user executes manually in TradeLocker. Faza 2 (auto-execution) deferred until prop firm TOS verified and Faza 1 proven over 20+ sessions.
-
-**Review changed two things from the original plan:**
-1. **State machine spec corrected.** Original "last 3 consecutive non-gray dots" is wrong. Actual M2D is phased: Phase 1 arming (turquoise → gray/dark-green) → Phase 2 trigger (light-green).
-2. **Levels extraction corrected.** Original plan had levels.py extracting SL/TP at trigger. But those lines only appear on TradeStation chart *after* user enters trade in TradeLocker. Corrected to two-phase: spec-math at trigger, chart-scan after entry.
-
-Plus 5 accepted expansions (labeled corpus, level fallback, layout canary, trade journal, TOS checklist).
-
----
-
-## Approach: B (Structured Python service, dry-run, audit log) + CEO-reviewed additions
-
-Runs on Windows machine alongside TradeStation. `mss` screenshots → ROI color-sample on M2D MAPS strip → phased state machine → Discord webhook + Telegram bot → JSONL audit + trade journal → dry-run replay against labeled corpus.
-
----
-
-## State Machine Spec (corrected + exhaustive)
-
-States:
-- `IDLE`
-- `ARMED_BUY` — turquoise seen
-- `PRIMED_BUY` — turquoise + at least one dark-green seen
-- `ARMED_SELL` — yellow seen
-- `PRIMED_SELL` — yellow + at least one dark-red seen
-
-**Default rule:** any (state, event) pair not listed below → stay in current state, no action, log as `noise`.
-
-Transitions — BUY side:
-
-| From | Event | To | Action |
-|------|-------|-----|--------|
-| IDLE | turquoise | ARMED_BUY | log arm_ts |
-| IDLE | yellow | ARMED_SELL | log arm_ts (sell) |
-| IDLE | dark-green / dark-red / light-green / light-red / gray | IDLE | noise (log phase-skip if light-green/light-red) |
-| ARMED_BUY | gray | ARMED_BUY | persist |
-| ARMED_BUY | turquoise | ARMED_BUY | refresh arm_ts |
-| ARMED_BUY | dark-green | PRIMED_BUY | log prime_ts |
-| ARMED_BUY | yellow | ARMED_SELL | opposite rearm |
-| ARMED_BUY | dark-red | ARMED_BUY | ignore (minority noise) |
-| ARMED_BUY | light-green | IDLE | **skip detected** — no FIRE, log phase_skip |
-| ARMED_BUY | light-red | IDLE | skip detected, log |
-| PRIMED_BUY | dark-green | PRIMED_BUY | accumulate |
-| PRIMED_BUY | dark-red | PRIMED_BUY | ignore (minority noise) |
-| PRIMED_BUY | **light-green** | IDLE | **FIRE BUY**, lockout(BUY)=4min |
-| PRIMED_BUY | light-red | IDLE | skip detected (wrong trigger) |
-| PRIMED_BUY | gray | IDLE | **COOLED** — signal dead, log |
-| PRIMED_BUY | turquoise | ARMED_BUY | rearm fresh |
-| PRIMED_BUY | yellow | ARMED_SELL | opposite rearm |
-
-SELL side mirrors exactly: swap turquoise↔yellow, dark-green↔dark-red, light-green↔light-red, BUY↔SELL.
-
-Notes:
-- No time-based TTL on ARMED/PRIMED. State persists until trigger fires, cooled by gray after PRIMED, opposite-color rearm, or process restart (Windows Task Scheduler stops bot at session end → natural session-boundary reset).
-- Cooling rule: "gray after dark-green" = signal racit (user's term). Gray during ARMED_BUY (before any dark-green) is OK.
-- After FIRE: 4-minute lockout per-direction. BUY lockout doesn't block SELL and vice versa. Single timestamp per direction.
-- Opposite-color-Phase-1 triggers rearm to opposite side (captures direction flip).
-- Phase-skip (arming color → trigger color with no phase-2 step) → IDLE, no FIRE, logged. Would be legitimate only if indicator collapses phases, which it doesn't per observed behavior.
-
----
-
-## Detection Details
-
-- **Loop interval:** 5 seconds (36 cycles per 3-min bar; stays well inside notification-latency target).
-- **Rightmost-dot detection:** scan ROI from right edge leftward, find first non-background pixel cluster → that's the rightmost dot. Don't hardcode x-pixel positions (chart scrolls; hardcoded positions drift).
-- **Debounce:** configurable `debounce_depth` in config.toml (default `1` — single-read acceptance). Increase if future sessions show mid-bar color flicker. Screenshot-in-notification is the user's visual verification on top.
-- **Rolling window:** keep last 20 classified dots with their detection timestamps. State machine consumes the newest *accepted* (post-debounce) dot per cycle.
-- **Classification:** nearest-color match in RGB Euclidean distance, per-color tolerance from calibration. Report confidence = `1 - distance_nearest / distance_second_nearest`. Log confidence every cycle. If all distances > tolerance → `UNKNOWN`, state unchanged.
-
----
-
-## Levels Extraction (two-phase, simplified)
-
-**Phase A — at trigger (immediate alert to Discord + Telegram):**
-- No entry-price compute. No spec-math SL/TP. User places a manual 0.6% SL in TradeLocker at entry; actual TP1/TP2/SL come in Phase B from the chart.
-- Notification: `🟢 BUY signal DIA→US30 | 22:47:03` + annotated screenshot (detected dot highlighted).
-
-**Phase B — after user trades (chart-scan confirmation):**
-- After Phase A fires, detector keeps watching the chart ROI for horizontal colored lines (red=SL, green=TP1/TP2).
-- When lines appear (user has entered trade in TradeLocker and TradeStation drew them) → scan y-pixels via Hough + color mask, convert via y-axis calibration → send second alert to both channels: `✅ Levels: SL=484.35 | TP1=485.20 | TP2=485.88`.
-- If chart-line scan times out (no lines in 10 min) → silent (user didn't trade).
-- If only 2 lines detected (user didn't set TP2 or line not rendered yet) → partial-result alert.
-- Phase B overlap with next signal: guarded by per-direction lockout + Phase-B completion flag; a new FIRE cannot issue until prior Phase B closes (timeout or success).
-
----
-
-## Dedup / Lockout
-
-- Time-based lockout: after any FIRE, block re-fire for 4 minutes (one 3-min bar + 1 min safety).
-- Tracked per-direction: BUY lockout doesn't block SELL.
-- Stored as single timestamp per direction (not pixel-keyed).
-
----
-
-## Observability
-
-- **Heartbeat:** every 30 min to a separate Discord thread (not main alerts channel): `🟢 22:00 alive | 0 triggers | confidence avg 0.85 | chart OK`. Silence >35 min = watchdog concern (user notices).
-- **Layout canary:** every 60 cycles (5 min), hash a stable reference region (axis labels, chart border). Stored baseline in config. On significant divergence (>threshold) → `⚠️ Layout changed — auto-paused, recalibrate` to alerts channel. Bot pauses detection until operator acknowledges (touch a pause-file or restart).
-- **Low-confidence alert:** 3+ consecutive cycles with confidence below threshold → `⚠️ Bot lost sight` (already in original plan).
-- **Window-lost alert:** TradeStation window not found for 60s → `⚠️ Cannot find chart`.
-- **Audit JSONL:** per-cycle, daily rotation (`logs/YYYY-MM-DD.jsonl`), fields: `{ts, window_found, roi_ok, rightmost_dot_color, confidence, state, transition, trigger, notified, reason}`.
-
----
-
-## Files to Create
-
-- `/workspace/atm/pyproject.toml` — Python 3.11+ required. Deps: `mss`, `opencv-python`, `numpy`, `requests`, `pygetwindow`, `pywin32` (DPI + window capture), `rich` (CLI), `pillow` (screenshot annotation). **No `tomli` — use stdlib `tomllib`.**
-- `/workspace/atm/config.toml` — populated by calibration tool (ROI coords, per-color RGB + tolerance, `debounce_depth`, y-axis scale, canary-region baseline hash, Discord webhook URL, Telegram bot token + chat_id)
-- `/workspace/atm/src/atm/config.py` — **[ENG-REVIEW]** `@dataclass Config` with `Config.load(path)` that validates on load (RGB tuples, positive tolerances, both notifier credentials present, y-axis 2-point pair). Fail fast at startup.
-- `/workspace/atm/src/atm/vision.py` — **[ENG-REVIEW]** shared primitives: ROI crop, perceptual hash, pixel-to-price linear interp, Hough line detection with color mask. Used by detector/canary/levels to avoid drift.
-- `/workspace/atm/src/atm/detector.py` — screenshot loop, rightmost-dot scan, color classification, rolling window, debounce
-- `/workspace/atm/src/atm/state_machine.py` — explicit phased state machine (spec above), exhaustive transition table
-- `/workspace/atm/src/atm/levels.py` — Phase B chart-scan only (Phase A entry-price compute removed after ENG-REVIEW)
-- `/workspace/atm/src/atm/canary.py` — layout fingerprint hash + drift check + auto-pause
-- `/workspace/atm/src/atm/notifier/__init__.py` — abstract `Notifier` protocol: `send_alert()`, `send_heartbeat()`, `send_levels_confirm()`
-- `/workspace/atm/src/atm/notifier/fanout.py` — **[ENG-REVIEW]** `FanoutNotifier` wraps N backends, each with its own worker thread + bounded queue (size 50, drop-oldest on overflow) + retry with exponential backoff + dead-letter file on total failure. Main loop never blocks.
-- `/workspace/atm/src/atm/notifier/discord.py` — webhook POST, annotated screenshot upload (multipart)
-- `/workspace/atm/src/atm/notifier/telegram.py` — **[ENG-REVIEW]** built in parallel with Discord (no longer deferred); bot API, photo upload
-- `/workspace/atm/src/atm/audit.py` — JSONL logger with daily local-midnight rotation, line-buffered write for crash safety
-- `/workspace/atm/src/atm/calibrate.py` — Tkinter: window pick → DPI check → ROI corners → per-color sample → y-axis scale → canary region → save versioned config
-- `/workspace/atm/src/atm/labeler.py` — **[EXPANSION]** Tkinter label UI → `labels.json`
-- `/workspace/atm/src/atm/dryrun.py` — replay with precision/recall/confusion matrix when labels present
-- `/workspace/atm/src/atm/journal.py` — **[EXPANSION]** `atm journal` CLI → `trades.jsonl`
-- `/workspace/atm/src/atm/report.py` — **[EXPANSION]** weekly aggregation
-- `/workspace/atm/src/atm/main.py` — CLI: `atm calibrate`, `atm label <dir>`, `atm dryrun <dir>`, `atm run [--duration Xh]`, `atm journal`, `atm report [--week YYYY-WW]`
-- `/workspace/atm/tests/` — **[ENG-REVIEW]** unit + E2E per test plan at `~/.gstack/projects/romfast-workspace/claude-master-eng-review-test-plan-20260415-212932.md`
-- `/workspace/atm/samples/`, `/workspace/atm/logs/`
-- `/workspace/atm/configs/` — versioned config archive. **[ENG-REVIEW]** No symlink (Windows admin-required); use `configs/current.txt` marker file storing the active filename. `Config.load()` reads the marker.
-- `/workspace/atm/docs/phase2-prop-firm-audit.md` — structured TOS checklist
-- `/workspace/atm/README.md` — setup, calibration workflow, per-session operating checklist, DPI/multi-monitor notes
-
----
-
-## Build Order
-
-1. **`pyproject.toml` + package scaffold** — Python 3.11+, `pip install -e .`, `atm --help` works.
-2. **Standalone screenshot-dump script** — `mss` timer dumps to `samples/` every 5s during trading sessions. Build corpus in parallel.
-3. **`config.py` + `vision.py`** — Config dataclass with validation; shared vision primitives. Ship with unit tests for config load + pixel-to-price interp.
-4. **`calibrate.py`** — versioned config in `configs/YYYY-MM-DD-HHMM.toml`; `configs/current.txt` marker file points at active. DPI check + canary region capture.
-5. **`labeler.py`** — once ~30 samples exist, tag them. `labels.json` is ground truth.
-6. **`state_machine.py`** + **unit tests** (clean BUY, clean SELL, cooling, opposite-rearm, lockout per-direction, noise, phase-skip, all state×color pairs via parameterized test).
-7. **`detector.py`** + **unit tests** (empty/background ROI, rightmost-cluster, rolling window FIFO, debounce depth=1, classification edges including UNKNOWN).
-8. **`canary.py`** + **unit tests** (drift threshold, pause-file gating).
-9. **`levels.py`** (Phase B only) + **unit tests** (Hough line detection with color mask, 2 vs 3 lines, 10-min timeout, pixel-to-price roundtrip).
-10. **`notifier/fanout.py` + `discord.py` + `telegram.py`** + **unit tests** (queue overflow drop-oldest, 429 backoff, dead-letter on total failure, fanout: one backend down still delivers). Both channels built in parallel — fire together from day 1.
-11. **`audit.py`** + **unit tests** (daily rotation at local midnight, line-buffered flush crash safety).
-12. **`dryrun.py`** — replay on `samples/` against `labels.json`. **Acceptance gate before live: precision = 100%, recall ≥ 95%.**
-13. **E2E replay test** — feed `samples/` through detector → state_machine → notifier-mock → in-memory audit; assert labels match FIREs.
-14. **`journal.py`**, **`report.py`**, **`main.py`** (unified CLI).
-15. **Windows Task Scheduler setup** — 16:30→18:30, 21:00→23:00. `atm run --duration 2h`. Manual DST check twice yearly.
-16. **`docs/phase2-prop-firm-audit.md`** — TOS checklist template.
-
----
-
-## Existing Utilities to Reuse
-
-Greenfield Python project. No internal utilities. External libs: `mss` (screenshot), `pygetwindow` (window locate), `opencv-python` (line detection in Phase B), `numpy` (color math), `requests` (Discord webhook), `tomli` (config parsing), `pillow` (annotated screenshots).
-
----
-
-## Verification
-
-End-to-end, in build order:
-
-1. **State machine unit tests:** `pytest tests/test_state_machine.py` — all scenarios (clean BUY, clean SELL, cooling, rearm, lockout, noise) pass.
-2. **Calibration:** `atm calibrate` → step through → `config.toml` populated with plausible RGBs for described colors + y-axis scale sane + canary region picked.
-3. **Labeled corpus:** ≥30 screenshots in `samples/`, `atm label ./samples` tags each.
-4. **Dry-run with metrics:** `atm dryrun ./samples` → precision + recall + confusion matrix printed. **Acceptance gate:** precision = 100%, recall ≥ 95%. If not met → tune tolerances, re-run.
-5. **Live test notification-only (2 sessions):** `atm run`. Verify:
-   - Discord + Telegram notifications within 5s of trigger, both channels receive.
-   - Phase A message: direction + timestamp + annotated screenshot.
-   - Phase B levels-alert fires once TradeStation draws SL/TP lines; correct SL/TP1/TP2 prices.
-   - Heartbeat messages every 30 min in thread.
-   - Audit JSONL complete, state transitions visible.
-   - Kill one notifier (e.g. wrong token) → other still delivers, dead-letter file for failed one.
-6. **Canary test:** manually move TradeStation window during session → layout-changed alert within 5 min. Move back → restart bot → resumes.
-7. **Scheduler test:** Windows Task Scheduler starts bot at 16:30, stops at 18:30 cleanly, log rotates at midnight.
-8. **Journal test:** after real trade, `atm journal` → prompt flow complete → `trades.jsonl` entry present.
-9. **Report test:** after 1 week of live use, `atm report --week 2026-16` → precision per color, slippage distribution, P&L summary.
-
----
-
-## Risk Register
-
-- **Prop firm TOS (Faza 2 blocker):** read TOS using `docs/phase2-prop-firm-audit.md` checklist before any auto-execution work. If EA/automation prohibited → Faza 2 dead, stay on Faza 1 permanently.
-- **TradeStation layout change:** canary catches it within 5 min → auto-pause. Recalibrate. Losing a session to a layout change is acceptable cost.
-- **Calibration drift over time:** versioned configs in `configs/` let you roll back to last-known-good if new calibration misfires.
-- **DIA↔US30 price divergence:** accepted (user's judgment). Phase 1 journal captures slippage per signal, feeding Faza 2 go/no-go.
-- **Screen sharing / RDP during trading:** overlay can break classification. Low prob, documented in README as operator hygiene.
-- **Windows Task Scheduler DST transitions:** twice per year, schedule may misfire. Manual check first week of each DST change.
-
----
-
-## Out of Scope (Faza 1)
-
-- Any automated click in TradeLocker (Faza 2 work)
-- Multi-symbol concurrent monitoring (single chart at a time; user switches manually between DIA and GLD)
-- Backtesting on historical data (strategy already manually validated)
-- Web UI / dashboard (headless + Discord/Telegram only)
-- Ack feedback loop (react-on-notification labeling) — deferred to TODOS.md as `P2-ack-loop`: shipping baseline first, adding feedback once detection quality verified
-- Telegram notifier — built only after Discord is stable 5+ sessions
-
----
-
-## Accepted Expansions (CEO review, SELECTIVE mode)
-
-1. ✅ **Labeled sample corpus + dry-run metrics** — `labeler.py`, `labels.json`, automated precision/recall in dryrun. Makes acceptance criteria ("false-positives = 0, false-negatives ≤ 5%") machine-checkable.
-2. ✅ **Level-extractor fallback (spec-math)** — Phase A always uses spec-math; Phase B validates against chart. Redundancy on fragile piece.
-3. ✅ **Layout canary + auto-pause** — `canary.py` hashes stable UI region, auto-pauses on drift. Catches silent classification-with-wrong-positions failure mode.
-4. ✅ **Trade journal CLI** — `atm journal` + `trades.jsonl` + weekly report. Data for Faza 2 go/no-go decision.
-5. ✅ **Prop-firm TOS audit checklist** — `docs/phase2-prop-firm-audit.md`. Structured Faza 2 evaluation framework shipped now.
-
-## Deferred to TODOS.md
-
-- **Ack feedback loop** — Discord reaction emojis feeding precision tuning. High value, operationally heavier (bot vs webhook). Add after Faza 1 baseline stable.
-
----
-
-## GSTACK REVIEW REPORT
-
-| Review | Trigger | Why | Runs | Status | Findings |
-|--------|---------|-----|------|--------|----------|
-| CEO Review | `/plan-ceo-review` | Scope & strategy | 1 | CLEAR (SELECTIVE EXPANSION) | 6 proposals, 5 accepted, 1 deferred; 2 arch corrections |
-| Codex Review | `/codex review` | Independent 2nd opinion | 0 | — | — |
-| Eng Review | `/plan-eng-review` | Architecture & tests (required) | 1 | CLEAR (FULL_REVIEW) | 9 issues found, 0 critical gaps; 4 decisions made, 0 unresolved |
-| Design Review | `/plan-design-review` | UI/UX gaps | 0 | — | SKIPPED (no UI scope — CLI + Discord/Telegram) |
-| DX Review | `/plan-devex-review` | Developer experience gaps | 0 | — | SKIPPED (personal tool, single user) |
-
-**UNRESOLVED:** 0
-
-**ENG REVIEW DECISIONS:**
-1. **Bar flicker** → debounce depth=1 (configurable), rely on screenshot-in-notification for visual verification.
-2. **Phase A entry price** → dropped. User places manual 0.6% SL in TradeLocker at entry. Phase A = direction + screenshot only. Phase B = real SL/TP1/TP2 from chart.
-3. **Notifier blocking** → fire-and-forget worker threads per backend, bounded queue (size 50, drop-oldest), retry w/ backoff, dead-letter on total failure.
-4. **Alert SPoF** → Discord + Telegram built in parallel from day 1, both fire together.
-
-**ENG REVIEW OBVIOUS FIXES (stated, no decision):**
-- Exhaustive state transition table (all state×color pairs, default-noise rule, SELL mirror explicit).
-- Python 3.11+ pin, drop `tomli` dep, use stdlib `tomllib`.
-- Windows symlink → `configs/current.txt` marker file.
-- Shared `vision.py` module (ROI, hash, interp, Hough).
-- `@dataclass Config` with fail-fast load-time validation.
-- DPI check + multi-monitor note in calibrate + README.
-
-**ENG REVIEW TEST SCOPE (accepted: FULL):** unit tests for every module (state_machine, detector, levels Phase B, canary, audit, notifier fanout/retry, calibrate roundtrip, config validate) + 1 E2E replay harness asserting labeled-corpus precision/recall. Test plan artifact: `~/.gstack/projects/romfast-workspace/claude-master-eng-review-test-plan-20260415-212932.md`.
-
-**VERDICT:** CEO + ENG CLEARED — ready to implement. Run `/ship` after implementation. No further reviews required before build.
diff --git a/docs/swirling-drifting-starfish.md b/docs/swirling-drifting-starfish.md
deleted file mode 100644
index 08b5643..0000000
--- a/docs/swirling-drifting-starfish.md
+++ /dev/null
@@ -1,74 +0,0 @@
-# Plan: ATM — Automated Trading Monitor (M2D, Faza 1)
-
-## Context
-
-User tranzacționează manual strategia M2D pe DIA (TradeStation) cu execuție pe TradeLocker US30 CFD (cont prop firm). Aceeași strategie merge și pe GLD → XAUUSD. 4 ore/seară trebuie să urmărească 2 ecrane. Obiectiv Faza 1: bot detectează automat trigger-ul și trimite notificare Telegram/Discord cu screenshot + nivele SL/TP1/TP2, user execută manual în TradeLocker. Faza 2 (auto-execution) deferată până prop firm TOS verificat + Faza 1 dovedită.
-
-Design doc complet salvat la `~/.gstack/projects/romfast-workspace/claude-master-design-20260415-atm-trading.md` (include strategia M2D cu toate detaliile).
-
-## Approach: B — Structured Python service + dry-run + audit log
-
-Rulează pe aceeași mașină Windows cu TradeStation. ROI color sampling pe strip-ul M2D MAPS, state machine pentru secvența 3-dot, notifier abstraction (Discord/Telegram), calibration Tkinter, dry-run pe screenshot-uri salvate.
-
-## Files to Create
-
-- `/workspace/atm/pyproject.toml` — packaging, deps: `mss`, `opencv-python`, `numpy`, `requests`, `pygetwindow`, `tomli`
-- `/workspace/atm/config.toml` — populat de calibration tool (ROI coords, culori referință + toleranțe, y-axis scale)
-- `/workspace/atm/src/atm/detector.py` — screenshot loop + color classification + state machine 3-dot
-- `/workspace/atm/src/atm/levels.py` — extragere SL/TP1/TP2 din liniile orizontale (pixel y → preț)
-- `/workspace/atm/src/atm/notifier/__init__.py` — interface `Notifier.send(signal, screenshot, levels)`
-- `/workspace/atm/src/atm/notifier/discord.py` — webhook POST
-- `/workspace/atm/src/atm/notifier/telegram.py` — bot API
-- `/workspace/atm/src/atm/audit.py` — JSONL logger, fiecare ciclu
-- `/workspace/atm/src/atm/calibrate.py` — Tkinter UI: click pe dot → capture RGB + tolerance; click pe colț ROI → salvează; click pe 2 puncte pe axa Y cu prețurile → calibrare scale
-- `/workspace/atm/src/atm/dryrun.py` — replay detector pe folder de screenshot-uri
-- `/workspace/atm/src/atm/main.py` — orchestration, CLI (`atm run`, `atm calibrate`, `atm dryrun <dir>`)
-- `/workspace/atm/samples/` — director screenshot-uri pentru dry-run corpus
-- `/workspace/atm/logs/` — director JSONL audit
-- `/workspace/atm/README.md` — setup + calibration workflow
-
-## Build Order
-
-1. **`pyproject.toml` + scaffold package** — `pip install -e .`, `atm --help` funcționează.
-2. **Script standalone de capture samples** (înainte de orice logică) — rulezi în timpul următoarelor sesiuni trading, dump screenshot la 5s interval în `samples/`. Ai corpus pentru dry-run.
-3. **`calibrate.py`** — fără config calibrat, nimic nu merge. Tkinter cu: pas 1 (select TradeStation window by title), pas 2 (click pe colțuri ROI M2D MAPS), pas 3 (click pe fiecare culoare: turquoise, verde închis, verde deschis, galben, roșu închis, roșu deschis + gri neutru; capturează RGB + rază de toleranță implicită 20), pas 4 (2 click-uri pe axa Y + valori preț introduse → scale factor pixel→preț). Salvează `config.toml`.
-4. **`detector.py`** — loop 1s: locate window, screenshot ROI, sample rightmost 5 dots pe pozițiile calibrate, clasifică fiecare la cea mai apropiată culoare (Euclidean in RGB cu toleranță). Rolling window ultimele 10 clasificări + timestamp. State machine: ultimele 3 non-gri consecutive = secvență BUY sau SELL? Fire o dată pe trigger (dedup set cu TTL 10min).
-5. **`levels.py`** — după trigger, scan chart region pentru liniile orizontale roșii (SL) și verzi (TP1/TP2). Extrage y-pixel al fiecărei linii, convertește la preț folosind scale-ul calibrat.
-6. **`notifier/discord.py`** — POST multipart cu screenshot adnotat + mesaj formatat: `🟢 BUY DIA→US30 | SL: 484.35 | TP1: 485.20 | TP2: 485.90 | 22:47:03`.
-7. **`dryrun.py`** — iterează `samples/`, rulează detector, printează ce AR fi trimis. Validare logică detecție înainte de live.
-8. **`audit.py`** — wrap detector loop, scrie JSONL: `{ts, window_found, roi_ok, dots:[...], classification:[...], trigger:null|"BUY"|"SELL", notified:true|false, reason}`.
-9. **`main.py`** — CLI unificat. `atm calibrate`, `atm dryrun ./samples`, `atm run` (loop live cu audit).
-10. **Windows Task Scheduler** — 2 task-uri: start 16:30 (stop 18:30), start 21:00 (stop 23:00). `atm run --duration 2h`.
-11. **`notifier/telegram.py`** — opțional după ce Discord e stabil.
-
-## Existing Utilities to Reuse
-
-N/A — greenfield project. No internal utilities to reuse.
-
-## Verification
-
-End-to-end, în ordinea din build:
-
-1. **Calibration workflow:** `atm calibrate` → urmezi pașii → rezultă `config.toml` complet. Verifică manual că RGB-urile sunt plauzibile pentru culorile descrise.
-2. **Dry-run corpus:** ai ≥20 screenshot-uri din sesiuni reale în `samples/`. Rulezi `atm dryrun ./samples` → output per screenshot: clasificare + decizie trigger. Manual verifici că cazurile unde ai văzut tu semnal reali → trigger; cazurile neutre → no-trigger. False-positives = 0 țintă, false-negatives ≤ 5%.
-3. **Live test notification-only (2 sesiuni):** `atm run` în fereastra trading. Verifici:
-   - Notificările Discord apar în 3s de când vezi trigger-ul pe chart.
-   - Screenshot atașat e clar, lizibil.
-   - SL/TP1/TP2 extrase sunt la ≤$0.05 de nivelele reale pe chart.
-   - Audit log (`logs/YYYY-MM-DD.jsonl`) conține fiecare ciclu; poți reproduce un missed signal.
-4. **Sanity alerts:** mută/redimensionează fereastra TradeStation → bot detectează "window lost" în 60s → notificare. Restabilește fereastra → bot reia.
-5. **Scheduler validation:** Windows Task Scheduler pornește `atm run` la 16:30, se oprește curat la 18:30, audit log salvează fără corupere.
-
-## Risk Register
-
-- **Prop firm TOS (Faza 2 blocker, NU Faza 1):** înainte de orice extensie spre auto-execution în TradeLocker, citești TOS-ul prop-ului, cauți "EA / automation / bot / copy trading / external signals". Dacă e interzis, Faza 2 e moartă și rămâi permanent pe Faza 1.
-- **Indicator layout change:** dacă TradeStation update schimbă render-ul M2D MAPS → re-calibration. Audit log va arăta degradare graduală a confidence-ului → alert activ via "bot lost sight".
-- **Price divergence DIA↔US30:** trigger-ul se dă pe DIA; poate fi o secundă unde US30 deja a mișcat diferit. Risc acceptabil (judgment user), dar monitorizat în Faza 2 prin slippage analysis.
-- **Screenshot pe ecran sharing / AnyDesk / RDP:** dacă cineva se conectează remote la Windows-ul tău în timpul trading, screenshot-urile pot cuprinde overlay-uri nepotrivite. Mic, dar notabil.
-
-## Out of Scope (Faza 1)
-
-- Orice click automat în TradeLocker
-- Multi-symbol concurrent monitoring (single chart la un moment dat)
-- Backtesting pe date istorice (strategia e deja validată manual)
-- UI / dashboard web — totul rulează headless cu notificări externe
diff --git a/logs/.gitkeep b/logs/.gitkeep
new file mode 100644
index 0000000..e69de29
diff --git a/pyproject.toml b/pyproject.toml
index 6f8a0a1..b6bb738 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -13,6 +13,7 @@ dependencies = [
     "pillow>=10.0",
     "requests>=2.31",
     "rich>=13.0",
+    "httpx>=0.27",
 ]
 
 [project.optional-dependencies]
@@ -24,6 +25,7 @@ windows = [
 dev = [
     "pytest>=8.0",
     "pytest-cov>=5.0",
+    "pytest-asyncio>=0.23",
 ]
 
 [project.scripts]
diff --git a/samples/calibration_labels.README.md b/samples/calibration_labels.README.md
new file mode 100644
index 0000000..216b9d0
--- /dev/null
+++ b/samples/calibration_labels.README.md
@@ -0,0 +1,33 @@
+# calibration_labels.json — schema
+
+Used by `atm validate-calibration` to check that the current color calibration
+classifies known-good screenshots correctly before a live session.
+
+## Schema
+
+A JSON array of entries. Each entry:
+
+| Field      | Type    | Required | Description                                                    |
+|------------|---------|----------|----------------------------------------------------------------|
+| `path`     | string  | yes      | Path to a PNG frame (relative to CWD or absolute).             |
+| `expected` | string  | yes      | Expected color name: one of `turquoise`, `yellow`, `dark_green`, `dark_red`, `light_green`, `light_red`, `gray`. |
+| `note`     | string  | no       | Freeform annotation; shown in SUGGESTIONS output.              |
+
+## Usage
+
+```bash
+atm validate-calibration samples/calibration_labels.json
+```
+
+Exit codes:
+- `0` — every sample PASS
+- `1` — one or more FAIL
+- `2` — label file missing or malformed JSON
+
+## Adding new samples
+
+1. Find a screenshot in `logs/fires/` whose dot color you can verify by eye.
+2. Append an entry with `path`, `expected`, and an optional `note`.
+3. Re-run validation. If it FAILs, the SUGGESTIONS section will tell you the
+   RGB distance between the observed pixel and the expected color's center —
+   use that as input for `atm calibrate`.
diff --git a/samples/calibration_labels.json b/samples/calibration_labels.json
new file mode 100644
index 0000000..031a0df
--- /dev/null
+++ b/samples/calibration_labels.json
@@ -0,0 +1,17 @@
+[
+  {
+    "path": "logs/fires/20260417_201500_arm_sell.png",
+    "expected": "yellow",
+    "note": "first arm of SELL cycle 2026-04-17"
+  },
+  {
+    "path": "logs/fires/20260417_205302_ss.png",
+    "expected": "dark_red",
+    "note": "user confirmed via screenshot (missed live alert)"
+  },
+  {
+    "path": "logs/fires/20260417_210441_ss.png",
+    "expected": "light_red",
+    "note": "fire phase (missed live alert)"
+  }
+]
diff --git a/src/atm/audit.py b/src/atm/audit.py
index 7a9e1a8..19d12ba 100644
--- a/src/atm/audit.py
+++ b/src/atm/audit.py
@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import json
+import threading
 from datetime import datetime, date
 from pathlib import Path
 from typing import Callable, IO
@@ -16,21 +17,25 @@ class AuditLog:
         self._clock: Callable[[], datetime] = clock or datetime.now
         self._current_date: date | None = None
         self._fh: IO[str] | None = None
+        self._lock = threading.Lock()
 
     def log(self, event: dict) -> None:
         now = self._clock()
         today = now.date()
-
-        if today != self._current_date:
-            self._open(today)
-
         if "ts" not in event:
             event = {**event, "ts": now.isoformat()}
-
-        assert self._fh is not None
-        self._fh.write(json.dumps(event, separators=(",", ":")) + "\n")
+        with self._lock:
+            if today != self._current_date:
+                self._open(today)
+            assert self._fh is not None
+            self._fh.write(json.dumps(event, separators=(",", ":")) + "\n")
 
     def close(self) -> None:
+        with self._lock:
+            self._close_locked()
+
+    def _close_locked(self) -> None:
+        """Close file handle; must be called while holding self._lock."""
         if self._fh is not None:
             try:
                 self._fh.close()
@@ -47,7 +52,7 @@ class AuditLog:
         return self._base_dir / f"{self._current_date}.jsonl"
 
     def _open(self, today: date) -> None:
-        self.close()
+        self._close_locked()  # already holding self._lock
         self._base_dir.mkdir(parents=True, exist_ok=True)
         path = self._base_dir / f"{today}.jsonl"
         self._fh = open(path, "a", buffering=1, encoding="utf-8")
diff --git a/src/atm/canary.py b/src/atm/canary.py
index 5dea2bd..602caf2 100644
--- a/src/atm/canary.py
+++ b/src/atm/canary.py
@@ -1,14 +1,18 @@
 """Layout drift detector via perceptual hash comparison."""
 from __future__ import annotations
 
+import logging
 from dataclasses import dataclass
 from pathlib import Path
+from typing import Callable
 
 import numpy as np
 
 from .config import Config
 from .vision import crop_roi, hamming_hex, phash
 
+logger = logging.getLogger(__name__)
+
 
 @dataclass
 class CanaryResult:
@@ -28,10 +32,15 @@ class Canary:
         self,
         cfg: Config,
         pause_flag_path: Path | None = None,
+        on_pause_callback: Callable[[int], None] | None = None,
     ) -> None:
         self._cfg = cfg
         self._pause_flag_path = pause_flag_path
         self._paused = False
+        # Single-shot callback invoked exactly once per not_paused→paused transition.
+        # Wrapped in try/except at call site so a faulty notifier never breaks
+        # the detection cycle.
+        self._on_pause = on_pause_callback
 
     def check(self, frame_bgr: np.ndarray) -> CanaryResult:
         roi_img = crop_roi(frame_bgr, self._cfg.canary.roi)
@@ -43,6 +52,12 @@ class Canary:
             self._paused = True
             if self._pause_flag_path is not None:
                 self._pause_flag_path.write_text("paused", encoding="utf-8")
+            if self._on_pause is not None:
+                try:
+                    self._on_pause(distance)
+                except Exception as exc:
+                    # Never let a notifier hiccup abort the detection cycle.
+                    logger.warning("canary on_pause_callback raised: %s", exc)
 
         return CanaryResult(distance=distance, drifted=drifted, paused=self._paused)
 
diff --git a/src/atm/commands.py b/src/atm/commands.py
new file mode 100644
index 0000000..2d6f6cf
--- /dev/null
+++ b/src/atm/commands.py
@@ -0,0 +1,170 @@
+"""Telegram command poller + Command dataclass.
+
+Uses httpx (async) for long-polling getUpdates. The sync TelegramNotifier
+continues to use requests — this module is the only httpx consumer.
+"""
+from __future__ import annotations
+
+import asyncio
+import logging
+from dataclasses import dataclass
+from typing import TYPE_CHECKING, Literal
+
+import httpx
+
+if TYPE_CHECKING:
+    from .config import TelegramCfg
+
+logger = logging.getLogger(__name__)
+
+CommandAction = Literal["set_interval", "stop", "status", "ss", "pause", "resume"]
+
+_BASE = "https://api.telegram.org/bot{token}/{method}"
+
+
+@dataclass
+class Command:
+    action: CommandAction
+    value: int | None = None  # seconds; only for set_interval
+
+
+class TelegramPoller:
+    """Long-poll Telegram getUpdates, emit Commands into asyncio.Queue.
+
+    Security: rejects messages from chat_ids not in cfg.allowed_chat_ids.
+    Degrades (stops polling) after 3 consecutive 401 responses and warns
+    via Discord (caller responsibility — poller only logs + sets degraded flag).
+    """
+
+    def __init__(
+        self,
+        cfg: TelegramCfg,
+        cmd_queue: asyncio.Queue[Command],
+        audit,  # _AuditLike
+    ) -> None:
+        self._cfg = cfg
+        self._cmd_queue = cmd_queue
+        self._audit = audit
+        self._offset = 0
+        self._consecutive_401 = 0
+        self._degraded = False
+        # fallback: if allowed_chat_ids is empty, accept only the primary chat
+        self._allowed = set(cfg.allowed_chat_ids) or {cfg.chat_id}
+
+    @property
+    def degraded(self) -> bool:
+        return self._degraded
+
+    async def run(self) -> None:
+        async with httpx.AsyncClient() as client:
+            await self._drain(client)
+            while True:
+                if self._degraded:
+                    await asyncio.sleep(5)
+                    continue
+                try:
+                    await self._poll_once(client)
+                except asyncio.CancelledError:
+                    raise
+                except (httpx.HTTPError, httpx.TimeoutException) as exc:
+                    self._audit.log({"event": "poller_error", "error": str(exc)})
+                    await asyncio.sleep(5)
+                except Exception as exc:  # json, unexpected
+                    self._audit.log({"event": "poller_error", "error": str(exc)})
+                    await asyncio.sleep(5)
+
+    async def _drain(self, client: httpx.AsyncClient) -> None:
+        """Discard all pending updates at startup so stale commands don't replay."""
+        try:
+            resp = await client.get(
+                _BASE.format(token=self._cfg.bot_token, method="getUpdates"),
+                params={"timeout": 0, "offset": self._offset},
+                timeout=10,
+            )
+            body = resp.json()
+            if body.get("ok") and body.get("result"):
+                self._offset = body["result"][-1]["update_id"] + 1
+        except Exception as exc:
+            logger.warning("TelegramPoller startup drain failed: %s", exc)
+
+    async def _poll_once(self, client: httpx.AsyncClient) -> None:
+        resp = await client.get(
+            _BASE.format(token=self._cfg.bot_token, method="getUpdates"),
+            params={"timeout": self._cfg.poll_timeout_s, "offset": self._offset},
+            timeout=self._cfg.poll_timeout_s + 5,
+        )
+
+        if resp.status_code == 401:
+            self._consecutive_401 += 1
+            if self._consecutive_401 >= 3:
+                self._degraded = True
+                self._audit.log({"event": "poller_degraded", "reason": "3_consecutive_401"})
+            return
+        self._consecutive_401 = 0
+
+        body = resp.json()
+        if not body.get("ok"):
+            return
+
+        for update in body.get("result", []):
+            self._offset = update["update_id"] + 1
+            await self._process_update(update)
+
+    async def _process_update(self, update: dict) -> None:
+        if "callback_query" in update:
+            # Inline button pressed — may be expired; reply with fallback
+            cbq = update["callback_query"]
+            chat_id = str(cbq.get("from", {}).get("id", ""))
+            if chat_id not in self._allowed:
+                logger.info("Rejected callback_query from chat_id=%s", chat_id)
+                return
+            # Caller handles answerCallbackQuery; just note in audit
+            self._audit.log({"event": "command_received", "action": "callback_query", "chat_id": chat_id})
+            return
+
+        msg = update.get("message") or update.get("edited_message")
+        if not msg:
+            return
+
+        chat_id = str(msg.get("chat", {}).get("id", ""))
+        if chat_id not in self._allowed:
+            logger.info("Rejected message from chat_id=%s", chat_id)
+            return
+
+        text = (msg.get("text") or "").strip().lower()
+        cmd = self._parse_command(text)
+        if cmd is None:
+            return
+
+        self._audit.log({
+            "event": "command_received",
+            "action": cmd.action,
+            "value": cmd.value,
+            "chat_id": chat_id,
+        })
+        await self._cmd_queue.put(cmd)
+
+    def _parse_command(self, text: str) -> Command | None:
+        t = text.lstrip("/").strip()
+        if not t:
+            return None
+        if t == "stop":
+            return Command(action="stop")
+        if t == "status":
+            return Command(action="status")
+        if t in ("ss", "screenshot"):
+            return Command(action="ss")
+        if t == "pause":
+            return Command(action="pause")
+        if t == "resume":
+            return Command(action="resume")
+        if t == "resume force":
+            # value=1 signals force: also lift canary drift-pause, not just user pause.
+            return Command(action="resume", value=1)
+        # "3" → set_interval 3 minutes → 180s; "interval 3" also accepted
+        parts = t.split()
+        if len(parts) == 1 and parts[0].isdigit():
+            return Command(action="set_interval", value=int(parts[0]) * 60)
+        if len(parts) == 2 and parts[0] in ("interval", "set_interval") and parts[1].isdigit():
+            return Command(action="set_interval", value=int(parts[1]) * 60)
+        return None
diff --git a/src/atm/config.py b/src/atm/config.py
index 4ff6ba2..6639304 100644
--- a/src/atm/config.py
+++ b/src/atm/config.py
@@ -5,6 +5,9 @@ import tomllib
 from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Literal
+from zoneinfo import ZoneInfo, ZoneInfoNotFoundError
+
+_VALID_WEEKDAYS: tuple[str, ...] = ("MON", "TUE", "WED", "THU", "FRI", "SAT", "SUN")
 
 DotColor = Literal[
     "turquoise", "yellow",
@@ -78,6 +81,9 @@ class DiscordCfg:
 class TelegramCfg:
     bot_token: str
     chat_id: str
+    allowed_chat_ids: tuple[str, ...] = ()
+    poll_timeout_s: int = 30
+    auto_poll_interval_s: int = 180
 
     def __post_init__(self) -> None:
         if not self.bot_token or not self.chat_id:
@@ -94,6 +100,43 @@ class AlertsCfg:
     trigger: bool = True
 
 
+@dataclass
+class OperatingHoursCfg:
+    """Session window: only run detection on allowed weekdays within HH:MM range.
+
+    Timezone is the source of truth for the exchange (default America/New_York
+    for NYSE). Start/stop are compared against the clock in that timezone.
+    Weekday check uses datetime.weekday() + a fixed MON..SUN list to stay
+    locale-independent (strftime('%a') returns localized names).
+
+    The ZoneInfo is cached at config load time so the detection loop doesn't
+    pay per-tick lookup cost.
+
+    NOTE: this dataclass is mutable (non-frozen) so Config._from_dict can stash
+    the resolved ZoneInfo onto `_tz_cache` after validation. Treat fields as
+    read-only at runtime.
+    """
+    enabled: bool = False
+    timezone: str = "America/New_York"
+    weekdays: tuple[str, ...] = ("MON", "TUE", "WED", "THU", "FRI")
+    start_hhmm: str = "09:30"
+    stop_hhmm: str = "16:00"
+    # Populated by Config._from_dict; None for disabled or failed-load cases.
+    _tz_cache: ZoneInfo | None = None
+
+
+@dataclass(frozen=True)
+class AlertBehaviorCfg:
+    """Alert behavior knobs (not screenshot toggles).
+
+    `fire_on_phase_skip`: backstop alert when FSM observes ARMED→light_{green,red}
+    directly (skipping the dark prime phase — often means dark color was
+    mis-classified as gray). Default True: missing a fire is worse than a noisy
+    phase-skip alert. Disable via `[options.alerts] fire_on_phase_skip = false`.
+    """
+    fire_on_phase_skip: bool = True
+
+
 @dataclass(frozen=True)
 class Config:
     window_title: str
@@ -114,6 +157,8 @@ class Config:
     phaseb_timeout_s: int = 600
     dead_letter_path: str = "logs/dead_letter.jsonl"
     attach_screenshots: AlertsCfg = field(default_factory=AlertsCfg)
+    alerts: AlertBehaviorCfg = field(default_factory=AlertBehaviorCfg)
+    operating_hours: OperatingHoursCfg = field(default_factory=OperatingHoursCfg)
     config_version: str = "unknown"
 
     def __post_init__(self) -> None:
@@ -156,9 +201,14 @@ class Config:
             drift_threshold=int(data["canary"].get("drift_threshold", 8)),
         )
         discord = DiscordCfg(webhook_url=data["discord"]["webhook_url"])
+        tg = data["telegram"]
+        _allowed = [str(c) for c in tg.get("allowed_chat_ids", [])] or [str(tg["chat_id"])]
         telegram = TelegramCfg(
-            bot_token=data["telegram"]["bot_token"],
-            chat_id=str(data["telegram"]["chat_id"]),
+            bot_token=tg["bot_token"],
+            chat_id=str(tg["chat_id"]),
+            allowed_chat_ids=tuple(_allowed),
+            poll_timeout_s=int(tg.get("poll_timeout_s", 30)),
+            auto_poll_interval_s=int(tg.get("auto_poll_interval_s", 180)),
         )
         opts = data.get("options", {})
         region = None
@@ -176,6 +226,36 @@ class Config:
             )
         else:
             attach = AlertsCfg()
+
+        alerts_dict = opts.get("alerts", {}) or {}
+        alert_behavior = AlertBehaviorCfg(
+            fire_on_phase_skip=bool(alerts_dict.get("fire_on_phase_skip", True)),
+        )
+
+        oh_dict = opts.get("operating_hours", {}) or {}
+        oh_weekdays = tuple(
+            str(w).upper() for w in oh_dict.get("weekdays", ("MON", "TUE", "WED", "THU", "FRI"))
+        )
+        for wd in oh_weekdays:
+            if wd not in _VALID_WEEKDAYS:
+                raise ValueError(
+                    f"operating_hours.weekdays contains invalid day {wd!r}; "
+                    f"expected any of {_VALID_WEEKDAYS}"
+                )
+        oh = OperatingHoursCfg(
+            enabled=bool(oh_dict.get("enabled", False)),
+            timezone=str(oh_dict.get("timezone", "America/New_York")),
+            weekdays=oh_weekdays,
+            start_hhmm=str(oh_dict.get("start_hhmm", "09:30")),
+            stop_hhmm=str(oh_dict.get("stop_hhmm", "16:00")),
+        )
+        if oh.enabled:
+            try:
+                oh._tz_cache = ZoneInfo(oh.timezone)
+            except ZoneInfoNotFoundError as exc:
+                raise ValueError(
+                    f"operating_hours.timezone {oh.timezone!r} invalid: {exc}"
+                ) from exc
         return cls(
             window_title=data["window_title"],
             dot_roi=roi,
@@ -195,5 +275,7 @@ class Config:
             phaseb_timeout_s=int(opts.get("phaseb_timeout_s", 600)),
             dead_letter_path=opts.get("dead_letter_path", "logs/dead_letter.jsonl"),
             attach_screenshots=attach,
+            alerts=alert_behavior,
+            operating_hours=oh,
             config_version=version,
         )
diff --git a/src/atm/detector.py b/src/atm/detector.py
index f2c3d82..e2bc603 100644
--- a/src/atm/detector.py
+++ b/src/atm/detector.py
@@ -28,6 +28,7 @@ class DetectionResult:
     match: ColorMatch | None  # None if no dot
     accepted: bool  # post-debounce; True only when match repeats debounce_depth times
     color: str | None  # accepted color name (UNKNOWN excluded)
+    dot_pos_abs: tuple[int, int] | None = None  # absolute (x, y) in frame; set when dot_found
 
 
 class Detector:
@@ -60,8 +61,14 @@ class Detector:
         self._debounce: deque[str | None] = deque(maxlen=cfg.debounce_depth)
         self._rolling: deque[DetectionResult] = deque(maxlen=20)
 
-    def step(self, ts: float) -> DetectionResult:
-        frame = self._capture()
+    def step(self, ts: float, frame=None) -> DetectionResult:
+        """Run one detection tick.
+
+        frame: pre-captured BGR ndarray (from asyncio.to_thread capture). When
+        None (default), calls self._capture() — preserving the sync-loop behaviour.
+        """
+        if frame is None:
+            frame = self._capture()
 
         if frame is None:
             self._debounce.append(None)
@@ -117,6 +124,7 @@ class Detector:
             match=match,
             accepted=accepted,
             color=color,
+            dot_pos_abs=(self._cfg.dot_roi.x + x, self._cfg.dot_roi.y + y),
         )
         self._rolling.append(r)
         return r
diff --git a/src/atm/main.py b/src/atm/main.py
index 4b71ff8..3c6589c 100644
--- a/src/atm/main.py
+++ b/src/atm/main.py
@@ -2,12 +2,15 @@
 from __future__ import annotations
 
 import argparse
+import asyncio
+import contextlib
 import os
 import sys
 import time
-from datetime import datetime
+from dataclasses import dataclass
+from datetime import datetime, tzinfo
 from pathlib import Path
-from typing import TYPE_CHECKING, Callable, Protocol, cast
+from typing import TYPE_CHECKING, Any, Callable, Protocol, cast
 
 from atm.config import Config  # stdlib-only (tomllib); safe at module level
 from atm.notifier import Alert
@@ -89,6 +92,23 @@ def main(argv=None) -> None:
         help="Stop at local HH:MM (overrides --duration). If the time is in "
              "the past when the loop starts, rolls over to tomorrow.",
     )
+    p_run.add_argument(
+        "--tz", metavar="ZONE", default=None,
+        help="Override operating_hours.timezone (e.g. America/New_York).",
+    )
+    p_run.add_argument(
+        "--weekdays", metavar="DAYS", default=None,
+        help="Override operating_hours.weekdays. Accepts comma list "
+             "(MON,TUE) or range (MON-FRI).",
+    )
+    p_run.add_argument(
+        "--oh-start", metavar="HH:MM", default=None,
+        help="Override operating_hours.start_hhmm (exchange-local).",
+    )
+    p_run.add_argument(
+        "--oh-stop", metavar="HH:MM", default=None,
+        help="Override operating_hours.stop_hhmm (exchange-local).",
+    )
 
     # journal
     p_journal = sub.add_parser("journal", help="Add a trade journal entry interactively")
@@ -115,6 +135,16 @@ def main(argv=None) -> None:
         metavar="PATH", help="Journal JSONL file (default: trades.jsonl)",
     )
 
+    # validate-calibration
+    p_valid = sub.add_parser(
+        "validate-calibration",
+        help="Offline: run Detector on labeled frames and report PASS/FAIL",
+    )
+    p_valid.add_argument(
+        "label_file", type=Path, metavar="LABEL_FILE",
+        help="JSON array with [{path, expected, note?}, ...] entries",
+    )
+
     args = parser.parse_args(argv)
 
     _dispatch = {
@@ -125,6 +155,7 @@ def main(argv=None) -> None:
         "debug": _cmd_debug,
         "journal": _cmd_journal,
         "report": _cmd_report,
+        "validate-calibration": _cmd_validate_calibration,
     }
     _dispatch[args.command](args)
 
@@ -168,6 +199,7 @@ def _cmd_dryrun(args) -> None:
 
 def _cmd_run(args) -> None:
     cfg = Config.load_current(Path("configs"))
+    cfg = _apply_operating_hours_cli_overrides(cfg, args)
     capture_stub = args.capture_stub or bool(os.environ.get("ATM_STUB_CAPTURE"))
 
     # --start-at HH:MM: sleep until the next occurrence of that local wall-clock time
@@ -227,6 +259,66 @@ def _cmd_run(args) -> None:
     run_live(cfg, duration_s=duration_s, capture_stub=capture_stub)
 
 
+_WEEKDAY_ORDER = ("MON", "TUE", "WED", "THU", "FRI", "SAT", "SUN")
+
+
+def _parse_weekdays_arg(raw: str) -> tuple[str, ...]:
+    """Accept 'MON,TUE,WED' or 'MON-FRI'. Case-insensitive."""
+    txt = raw.strip().upper()
+    if "-" in txt and "," not in txt:
+        a, b = (p.strip() for p in txt.split("-", 1))
+        if a not in _WEEKDAY_ORDER or b not in _WEEKDAY_ORDER:
+            raise ValueError(f"unknown weekday(s) in range {raw!r}")
+        i, j = _WEEKDAY_ORDER.index(a), _WEEKDAY_ORDER.index(b)
+        if i > j:
+            raise ValueError(f"weekday range reversed: {raw!r}")
+        return tuple(_WEEKDAY_ORDER[i : j + 1])
+    days = tuple(d.strip() for d in txt.split(",") if d.strip())
+    for d in days:
+        if d not in _WEEKDAY_ORDER:
+            raise ValueError(f"unknown weekday {d!r} (valid: {_WEEKDAY_ORDER})")
+    return days
+
+
+def _apply_operating_hours_cli_overrides(cfg, args):
+    """Return cfg (possibly new) with operating_hours overridden by CLI flags.
+
+    Config is a frozen dataclass, but operating_hours is non-frozen by design
+    so we can tweak it in-place and recompute the tz cache. CLI flags implicitly
+    enable operating_hours even if the TOML had it disabled.
+    """
+    import dataclasses as _dc
+    from zoneinfo import ZoneInfo, ZoneInfoNotFoundError
+
+    oh = cfg.operating_hours
+    any_override = any(
+        getattr(args, k, None)
+        for k in ("tz", "weekdays", "oh_start", "oh_stop")
+    )
+    if not any_override:
+        return cfg
+
+    new_tz = args.tz if args.tz else oh.timezone
+    try:
+        tz_cache = ZoneInfo(new_tz)
+    except ZoneInfoNotFoundError as exc:
+        sys.exit(f"--tz {new_tz!r} invalid: {exc}")
+
+    new_weekdays = _parse_weekdays_arg(args.weekdays) if args.weekdays else oh.weekdays
+    new_start = args.oh_start if args.oh_start else oh.start_hhmm
+    new_stop = args.oh_stop if args.oh_stop else oh.stop_hhmm
+    oh.enabled = True
+    oh.timezone = new_tz
+    oh.weekdays = new_weekdays
+    oh.start_hhmm = new_start
+    oh.stop_hhmm = new_stop
+    oh._tz_cache = tz_cache
+    # Config is frozen but operating_hours is a mutable field object —
+    # mutating it in place is sufficient; no dataclasses.replace needed.
+    _ = _dc  # keep import for future use
+    return cfg
+
+
 def _cmd_journal(args) -> None:
     try:
         from atm.journal import Journal, prompt_entry
@@ -337,6 +429,37 @@ def _cmd_report(args) -> None:
     )
 
 
+def _cmd_validate_calibration(args) -> None:
+    """Run offline calibration validation; exit 0 on 100% PASS, 1 otherwise."""
+    try:
+        from atm.validate import validate_calibration, ValidationError
+    except ImportError as exc:
+        sys.exit(f"validate module not available: {exc}")
+
+    label_file = Path(args.label_file)
+    try:
+        cfg = Config.load_current(Path("configs"))
+    except FileNotFoundError as exc:
+        sys.exit(f"config not found: {exc}")
+
+    try:
+        config_name = ""
+        cur_ptr = Path("configs") / "current.txt"
+        if cur_ptr.exists():
+            config_name = cur_ptr.read_text(encoding="utf-8").strip()
+    except Exception:
+        config_name = ""
+
+    try:
+        report = validate_calibration(label_file, cfg, config_name=config_name)
+    except ValidationError as exc:
+        print(f"error: {exc}", file=sys.stderr)
+        sys.exit(2)
+
+    print(report.render())
+    sys.exit(0 if report.all_pass else 1)
+
+
 # ---------------------------------------------------------------------------
 # Live loop
 # ---------------------------------------------------------------------------
@@ -348,6 +471,8 @@ def _save_annotated_frame(
     label: str,
     now: float,
     audit: _AuditLike | None = None,
+    dot_pos_abs: "tuple[int, int] | None" = None,
+    canary_ok: bool = True,
 ) -> "Path | None":
     """Save BGR frame with cyan dot_roi rect to ``logs/fires/{ts}_{label}.png``.
 
@@ -355,6 +480,10 @@ def _save_annotated_frame(
     audit (when provided) so disk-full / permission issues don't become silent
     regressions. Never raises — snapshot is a best-effort enhancement, the
     text alert must still go out.
+
+    dot_pos_abs + canary_ok: when both are set the price overlay is drawn
+    (y-axis linear interpolation via cfg.y_axis). Skipped when canary drifted
+    since calibration may be stale.
     """
     try:
         import cv2  # type: ignore[import-untyped]
@@ -371,6 +500,22 @@ def _save_annotated_frame(
         annotated = frame.copy()
         x, y, w, h = cfg.dot_roi.x, cfg.dot_roi.y, cfg.dot_roi.w, cfg.dot_roi.h
         cv2.rectangle(annotated, (x, y), (x + w, y + h), (0, 255, 255), 2)
+        if dot_pos_abs is not None and canary_ok and hasattr(cfg, "y_axis"):
+            try:
+                _, dot_y = dot_pos_abs
+                ya = cfg.y_axis
+                slope = (ya.p2_price - ya.p1_price) / (ya.p2_y - ya.p1_y)
+                price = ya.p1_price + (dot_y - ya.p1_y) * slope
+                w_frame = annotated.shape[1]
+                text = f"${price:.2f}"
+                font = cv2.FONT_HERSHEY_SIMPLEX
+                scale, thickness = 1.2, 3
+                (tw, th), _ = cv2.getTextSize(text, font, scale, thickness)
+                tx, ty = w_frame - tw - 10, th + 10
+                cv2.rectangle(annotated, (tx - 4, 4), (tx + tw + 4, ty + 4), (0, 0, 0), -1)
+                cv2.putText(annotated, text, (tx, ty), font, scale, (255, 255, 255), thickness, cv2.LINE_AA)
+            except Exception:
+                pass  # price overlay is best-effort; never break the screenshot
         cv2.imwrite(str(fpath), annotated)
         return fpath
     except Exception as exc:
@@ -390,6 +535,7 @@ def _handle_tick(
     audit: _AuditLike,
     first_accepted: bool,
     snapshot: Snapshot | None = None,
+    cfg: Any = None,
 ) -> Transition | None:
     """Feed FSM for a single accepted color and dispatch arm/prime/late_start
     alerts. Returns the final Transition, or None when the color triggered a
@@ -493,10 +639,464 @@ def _handle_tick(
             image_path=snap(prime_kind, prime_label),
             direction=direction,
         ))
+    # PHASE_SKIP fire backstop: ARMED→light_{green,red} directly (dark was missed).
+    # Emits a fire-equivalent alert when cfg.alerts.fire_on_phase_skip (default True).
+    # Uses public FSM lockout API (is_locked/record_fire) to reuse the standard
+    # 240s dedupe window so bouncing detectors do not spam the user.
+    elif tr.reason == "phase_skip" and color in ("light_green", "light_red"):
+        flag_on = True
+        if cfg is not None:
+            alerts_cfg = getattr(cfg, "alerts", None)
+            if alerts_cfg is not None:
+                flag_on = bool(getattr(alerts_cfg, "fire_on_phase_skip", True))
+        if flag_on:
+            direction = "BUY" if color == "light_green" else "SELL"
+            if not fsm.is_locked(direction, now):
+                fsm.record_fire(direction, now)
+                dark_name = "dark_green" if direction == "BUY" else "dark_red"
+                notifier.send(Alert(
+                    kind="phase_skip_fire",
+                    title=f"PHASE SKIP {direction} — {dark_name} nu a fost detectat",
+                    body=(
+                        "Verifică chart-ul manual. Posibil necalibrare culoare "
+                        f"(observat {color} direct după armare)."
+                    ),
+                    image_path=snap("phase_skip", f"phase_skip_{direction.lower()}"),
+                    direction=direction,
+                ))
     return tr
 
 
+@dataclass
+class _TickSyncResult:
+    frame: Any = None
+    res: Any = None       # DetectionResult | None
+    tr: Any = None        # Transition | None
+    first_consumed: bool = False
+    late_start: bool = False
+    new_color: str | None = None  # corpus sample color when changed
+
+
+@dataclass
+class RunContext:
+    """Dependencies passed to module-scope detection-loop helpers.
+
+    Keeps `_run_tick`, `_handle_fsm_result`, `_drain_cmd_queue`, and
+    `_dispatch_command` at module scope so they are directly unit-testable
+    without reconstructing `run_live_async`.
+    """
+    cfg: Any
+    capture: Callable
+    canary: Any
+    detector: Any
+    fsm: Any
+    notifier: _NotifierLike
+    audit: _AuditLike
+    detection_log: _AuditLike
+    scheduler: Any
+    samples_dir: Path
+    fires_dir: Path
+    cmd_queue: Any  # asyncio.Queue[Command]
+    state: Any      # carries first_accepted, last_saved_color, levels_extractor, fire_count, start
+    levels_extractor_factory: Callable  # builds LevelsExtractor(cfg, trigger, now)
+    lifecycle: Any = None  # LifecycleState — window + user_paused tracking
+
+
+@dataclass
+class _LoopState:
+    """Per-loop mutable state (previously closure nonlocals)."""
+    first_accepted: bool = True
+    last_saved_color: str | None = None
+    levels_extractor: Any = None
+    fire_count: int = 0
+    start: float = 0.0
+
+
+@dataclass
+class LifecycleState:
+    """Tracks user-pause / out-of-window state across detection ticks.
+
+    last_window_state: None at startup so _maybe_log_transition can seed it
+    without emitting a spurious market_open alert on the first in-window tick.
+    """
+    user_paused: bool = False
+    last_window_state: str | None = None  # "open" / "closed" / None (uninitialized)
+
+
+# Locale-independent weekday names; index matches datetime.weekday() (MON=0).
+_WEEKDAY_NAMES: tuple[str, ...] = ("MON", "TUE", "WED", "THU", "FRI", "SAT", "SUN")
+
+
+def _should_skip(now_ts: float, state: LifecycleState, cfg, canary) -> str | None:
+    """Return a reason string if detection should be skipped, else None.
+
+    Order: user_paused > canary drift > operating-hours window. Uses the
+    ZoneInfo cached on cfg.operating_hours._tz_cache (populated at config load)
+    to avoid per-tick tz lookup cost.
+    """
+    if state.user_paused:
+        return "user_paused"
+    if getattr(canary, "is_paused", False):
+        return "drift_paused"
+    oh = getattr(cfg, "operating_hours", None)
+    if oh is None or not oh.enabled:
+        return None
+    tz = getattr(oh, "_tz_cache", None)
+    if not isinstance(tz, tzinfo):
+        # Enabled but no tz resolved (or mock cfg in tests) — skip rather than crash.
+        return None
+    now_exchange = datetime.fromtimestamp(now_ts, tz=tz)
+    # weekday() = 0..6 (MON..SUN). Locale-free; strftime('%a') is not.
+    if _WEEKDAY_NAMES[now_exchange.weekday()] not in oh.weekdays:
+        return "out_of_window_weekend"
+    hhmm = now_exchange.strftime("%H:%M")
+    if hhmm < oh.start_hhmm or hhmm >= oh.stop_hhmm:
+        return "out_of_window_hours"
+    return None
+
+
+def _maybe_log_transition(
+    reason: str | None,
+    state: LifecycleState,
+    now: float,
+    audit: _AuditLike,
+    notifier: _NotifierLike,
+) -> None:
+    """Log market_open / market_closed exactly once per transition.
+
+    Startup guard (R2): when last_window_state is None we just seed it; no
+    alert/audit event is emitted for the initial evaluation. This prevents a
+    spurious market_open alert when run_live_async starts in-window.
+    """
+    if reason is None:
+        window_reason = "open"
+    elif reason.startswith("out_of_window"):
+        window_reason = "closed"
+    else:
+        # user_paused / drift_paused don't change market window state
+        return
+
+    if window_reason == state.last_window_state:
+        return
+
+    if state.last_window_state is None:
+        state.last_window_state = window_reason
+        return
+
+    event_name = "market_open" if window_reason == "open" else "market_closed"
+    audit.log({"ts": now, "event": event_name, "reason": reason})
+    body = (
+        "Piața închisă — monitorizare pauzată până la următoarea deschidere"
+        if event_name == "market_closed"
+        else "Piața deschisă — monitorizare reluată"
+    )
+    notifier.send(Alert(
+        kind="status",
+        title=event_name.replace("_", " ").title(),
+        body=body,
+    ))
+    state.last_window_state = window_reason
+
+
+def _sync_detection_tick(
+    capture: Callable,
+    canary: Any,
+    cfg: Any,
+    detector: Any,
+    fsm: Any,
+    notifier: _NotifierLike,
+    audit: _AuditLike,
+    detection_log: _AuditLike,
+    fires_dir: Path,
+    first_accepted: bool,
+    last_saved_color: "str | None",
+    now: float,
+    samples_dir: Path,
+) -> _TickSyncResult:
+    """One full detection tick (blocking I/O). Runs in asyncio.to_thread."""
+    frame = capture()
+    if frame is None:
+        audit.log({"ts": now, "event": "window_lost"})
+        return _TickSyncResult()
+
+    cr = canary.check(frame)
+    if canary.is_paused:
+        audit.log({"ts": now, "event": "paused", "drift": cr.distance})
+        return _TickSyncResult(frame=frame)
+
+    res = detector.step(now, frame)
+    detection_log.log({
+        "ts": now, "event": "frame",
+        "window_found": res.window_found,
+        "dot_found": res.dot_found,
+        "rgb": list(res.rgb) if res.rgb is not None else None,
+        "match_name": res.match.name if res.match is not None else None,
+        "distance": round(res.match.distance, 2) if res.match is not None else None,
+        "confidence": round(res.match.confidence, 3) if res.match is not None else None,
+        "accepted": res.accepted,
+        "color": res.color,
+    })
+
+    if not (res.accepted and res.color):
+        return _TickSyncResult(frame=frame, res=res)
+
+    is_first = first_accepted
+
+    def _snapshot(kind: str, label: str) -> "Path | None":
+        if not getattr(cfg.attach_screenshots, kind, True):
+            return None
+        return _save_annotated_frame(
+            frame, cfg, fires_dir, label, now, audit=audit,
+            dot_pos_abs=getattr(res, "dot_pos_abs", None),
+            canary_ok=True,
+        )
+
+    tr = _handle_tick(fsm, res.color, now, notifier, audit, is_first, snapshot=_snapshot, cfg=cfg)
+
+    if tr is None:
+        return _TickSyncResult(frame=frame, res=res, first_consumed=is_first, late_start=True)
+
+    new_color: str | None = None
+    if res.color != last_saved_color:
+        ts_str = datetime.fromtimestamp(now).strftime("%Y%m%d_%H%M%S")
+        sample_path = samples_dir / f"{ts_str}_{res.color}.png"
+        try:
+            import cv2  # type: ignore[import-untyped]
+            cv2.imwrite(str(sample_path), frame)
+        except Exception as exc:
+            audit.log({"ts": now, "event": "sample_save_failed", "error": str(exc)})
+        new_color = res.color
+
+    if tr.trigger and not tr.locked:
+        fire_path: "Path | None" = None
+        if cfg.attach_screenshots.trigger:
+            fire_path = _save_annotated_frame(
+                frame, cfg, fires_dir, tr.trigger, now, audit=audit,
+                dot_pos_abs=getattr(res, "dot_pos_abs", None),
+                canary_ok=True,
+            )
+        notifier.send(Alert(
+            kind="trigger",
+            title=f"Semnal {tr.trigger}",
+            body=f"@ {datetime.fromtimestamp(now).isoformat(timespec='seconds')}",
+            image_path=fire_path,
+            direction=tr.trigger,
+        ))
+
+    return _TickSyncResult(
+        frame=frame, res=res, tr=tr,
+        first_consumed=is_first, new_color=new_color,
+    )
+
+
+async def _run_tick(ctx: RunContext) -> _TickSyncResult:
+    """Execute one `_sync_detection_tick` in a thread; returns result or empty.
+
+    Lifecycle gating (user pause / operating hours / drift) happens here, not
+    inside the sync tick, so the async loop can still drain commands and emit
+    market_open / market_closed transitions even when the heavy detection
+    work is skipped.
+    """
+    now = time.time()
+    if ctx.lifecycle is not None:
+        skip = _should_skip(now, ctx.lifecycle, ctx.cfg, ctx.canary)
+        _maybe_log_transition(skip, ctx.lifecycle, now, ctx.audit, ctx.notifier)
+        if skip is not None:
+            # No detection this tick. Empty result → _handle_fsm_result no-op.
+            return _TickSyncResult()
+    return await asyncio.to_thread(
+        _sync_detection_tick,
+        ctx.capture, ctx.canary, ctx.cfg, ctx.detector, ctx.fsm,
+        ctx.notifier, ctx.audit, ctx.detection_log,
+        ctx.fires_dir, ctx.state.first_accepted, ctx.state.last_saved_color,
+        now, ctx.samples_dir,
+    )
+
+
+async def _handle_fsm_result(ctx: RunContext, result: _TickSyncResult) -> None:
+    """Scheduler start/stop + levels extraction. No-op if res is None/late_start."""
+    if result.first_consumed:
+        ctx.state.first_accepted = False
+    if result.new_color is not None:
+        ctx.state.last_saved_color = result.new_color
+
+    tr = result.tr
+    res = result.res
+
+    if result.late_start or res is None:
+        return
+
+    if tr is not None and getattr(res, "accepted", False) and getattr(res, "color", None):
+        if tr.reason == "prime" and not ctx.scheduler.is_running:
+            ctx.scheduler.start(ctx.cfg.telegram.auto_poll_interval_s)
+            ctx.audit.log({"ts": time.time(), "event": "scheduler_started", "reason": "primed"})
+        elif tr.reason in ("fire", "cooled", "phase_skip", "opposite_rearm") and ctx.scheduler.is_running:
+            ctx.scheduler.stop()
+            ctx.audit.log({"ts": time.time(), "event": "scheduler_stopped", "reason": tr.reason})
+
+    if tr is not None and tr.trigger and not tr.locked:
+        ctx.state.fire_count += 1
+        if ctx.scheduler.is_running:
+            ctx.scheduler.stop()
+            ctx.audit.log({"ts": time.time(), "event": "scheduler_stopped", "reason": "fire"})
+        ctx.state.levels_extractor = ctx.levels_extractor_factory(ctx.cfg, tr.trigger, time.time())
+
+    if ctx.state.levels_extractor is not None and result.frame is not None:
+        lr = ctx.state.levels_extractor.step(result.frame, time.time())
+        if lr.status in ("complete", "timeout"):
+            if lr.status == "complete" and lr.levels:
+                ctx.notifier.send(Alert(
+                    kind="levels",
+                    title="Niveluri",
+                    body=(
+                        f"SL={lr.levels.sl} "
+                        f"TP1={lr.levels.tp1} "
+                        f"TP2={lr.levels.tp2}"
+                    ),
+                ))
+            ctx.state.levels_extractor = None
+
+
+async def _dispatch_command(ctx: RunContext, cmd) -> None:
+    """Process a single Command. Exceptions bubble — caller wraps in try/except."""
+    cfg = ctx.cfg
+    if cmd.action == "set_interval":
+        secs = cmd.value or cfg.telegram.auto_poll_interval_s
+        ctx.scheduler.start(secs)
+        ctx.audit.log({"ts": time.time(), "event": "scheduler_started", "reason": "set_interval", "interval_s": secs})
+        ctx.notifier.send(Alert(kind="status", title=f"Polling activ — interval {secs // 60} min", body=""))
+    elif cmd.action == "stop":
+        if ctx.scheduler.is_running:
+            ctx.scheduler.stop()
+            ctx.audit.log({"ts": time.time(), "event": "scheduler_stopped", "reason": "command_stop"})
+            ctx.notifier.send(Alert(kind="status", title="Polling oprit", body=""))
+        else:
+            ctx.notifier.send(Alert(kind="status", title="Polling nu este activ", body=""))
+    elif cmd.action == "status":
+        uptime_s = time.monotonic() - ctx.state.start
+        last_roll = ctx.detector.rolling[-1] if ctx.detector.rolling else None
+        last_conf = f"{last_roll.match.confidence:.2f}" if last_roll and last_roll.match else "—"
+        last_color = (
+            (last_roll.color or last_roll.match.name) if last_roll and last_roll.match else "—"
+        ) if last_roll else "—"
+        sched_info = (
+            f"activ @{ctx.scheduler.interval_s // 60}min" if ctx.scheduler.interval_s else "activ"
+        ) if ctx.scheduler.is_running else "oprit"
+        canary_info = "drift (pauze)" if ctx.canary.is_paused else "ok"
+
+        # Active / pause reason + window state
+        active_info = "activ"
+        window_info = "—"
+        if ctx.lifecycle is not None:
+            skip = _should_skip(time.time(), ctx.lifecycle, ctx.cfg, ctx.canary)
+            if skip is not None:
+                active_info = f"pauzat:{skip}"
+            oh = getattr(ctx.cfg, "operating_hours", None)
+            if oh is not None and oh.enabled:
+                window_info = ctx.lifecycle.last_window_state or "—"
+            else:
+                window_info = "always_on"
+
+        body = (
+            f"Stare: {ctx.fsm.state.value}\n"
+            f"Activ: {active_info} | Fereastră: {window_info}\n"
+            f"Ultima detecție: {last_color} (conf {last_conf})\n"
+            f"Uptime: {uptime_s / 3600:.1f}h | Semnale: {ctx.state.fire_count}\n"
+            f"Poller: {sched_info} | Canary: {canary_info}"
+        )
+        ctx.notifier.send(Alert(kind="status", title="ATM Status", body=body))
+    elif cmd.action == "ss":
+        now_ss = time.time()
+        frame_ss = await asyncio.to_thread(ctx.capture)
+        if frame_ss is None:
+            ctx.notifier.send(Alert(
+                kind="warn",
+                title="Captură eșuată — verificați fereastra TradeStation",
+                body="",
+            ))
+            return
+        path_ss = await asyncio.to_thread(
+            _save_annotated_frame, frame_ss, ctx.cfg, ctx.fires_dir, "ss", now_ss, ctx.audit,
+        )
+        ctx.audit.log({"ts": now_ss, "event": "screenshot_sent", "path": str(path_ss) if path_ss else None})
+        ctx.notifier.send(Alert(kind="screenshot", title="Screenshot manual", body="", image_path=path_ss))
+    elif cmd.action == "pause":
+        # User manually stops monitoring. Canary drift state is untouched.
+        if ctx.lifecycle is not None:
+            ctx.lifecycle.user_paused = True
+        ctx.audit.log({"ts": time.time(), "event": "user_paused"})
+        ctx.notifier.send(Alert(
+            kind="status",
+            title="Monitorizare oprită manual",
+            body="Folosește /resume pentru a relua.",
+        ))
+    elif cmd.action == "resume":
+        # R2: /resume clears only user_paused. Canary drift requires
+        # /resume force (value == 1) so the user acknowledges the risk.
+        was_drift = bool(getattr(ctx.canary, "is_paused", False))
+        was_user = bool(ctx.lifecycle.user_paused) if ctx.lifecycle is not None else False
+        force = cmd.value == 1
+        if ctx.lifecycle is not None:
+            ctx.lifecycle.user_paused = False
+        if force and was_drift:
+            ctx.canary.resume()
+        ctx.audit.log({
+            "ts": time.time(), "event": "user_resumed",
+            "was_drift": was_drift, "was_user": was_user, "force": force,
+        })
+        # Adaptive response
+        if was_drift and not force:
+            title = "Pauză user eliminată — dar Canary drift activ"
+            body = (
+                "Trimite /resume force pentru a anula drift-pause. "
+                "Recalibrează dacă driftul persistă."
+            )
+        elif force and was_drift:
+            title = "Drift-pause anulat manual (force)"
+            body = "Dacă driftul persistă, Canary va repauza."
+        else:
+            skip_now = None
+            if ctx.lifecycle is not None:
+                skip_now = _should_skip(time.time(), ctx.lifecycle, ctx.cfg, ctx.canary)
+            if skip_now and skip_now.startswith("out_of_window"):
+                title = "Pauză eliminată — piața e închisă acum"
+                body = "Monitorizarea va porni la următoarea fereastră."
+            else:
+                title = "Monitorizare reluată"
+                body = ""
+        ctx.notifier.send(Alert(kind="status", title=title, body=body))
+
+
+async def _drain_cmd_queue(ctx: RunContext) -> None:
+    """Drain all pending commands, isolating each dispatch in try/except.
+
+    CRITICAL: this MUST run every loop iteration, unconditionally, even when
+    the detection tick returned nothing (canary paused, out-of-window, etc.).
+    Prior bug: the main loop `continue`'d past this drain when res=None,
+    causing commands to accumulate indefinitely while canary was drifted.
+    """
+    while True:
+        try:
+            cmd = ctx.cmd_queue.get_nowait()
+        except asyncio.QueueEmpty:
+            return
+        try:
+            await _dispatch_command(ctx, cmd)
+        except Exception as exc:
+            ctx.audit.log({
+                "ts": time.time(), "event": "command_error",
+                "action": cmd.action, "error": str(exc),
+            })
+            print(f"ERR command_dispatch /{cmd.action}: {exc}", flush=True)
+            ctx.notifier.send(Alert(kind="warn", title=f"Eroare comandă /{cmd.action}", body=str(exc)))
+
+
 def run_live(cfg, duration_s=None, capture_stub: bool = False) -> None:
+    """Sync entry point — delegates to asyncio event loop."""
+    asyncio.run(run_live_async(cfg, duration_s=duration_s, capture_stub=capture_stub))
+
+
+async def run_live_async(cfg, duration_s=None, capture_stub: bool = False) -> None:
     """Main live monitoring loop. Imports are lazy to keep --help fast."""
     try:
         from atm.detector import Detector
@@ -506,14 +1106,38 @@ def run_live(cfg, duration_s=None, capture_stub: bool = False) -> None:
         from atm.notifier.discord import DiscordNotifier
         from atm.notifier.telegram import TelegramNotifier
         from atm.audit import AuditLog
+        from atm.commands import TelegramPoller, Command
+        from atm.scheduler import ScreenshotScheduler
     except ImportError as exc:
         sys.exit(f"run-loop dependencies not available: {exc}")
 
     capture = _build_capture(cfg, capture_stub=capture_stub)
     detector = Detector(cfg, capture)
     fsm = StateMachine(lockout_s=cfg.lockout_s)
-    canary = Canary(cfg, pause_flag_path=Path("logs/pause.flag"))
     audit = AuditLog(Path("logs"))
+
+    # Forward-declare notifier so the canary pause callback can close over it.
+    # The notifier is constructed a few lines below once backends exist.
+    _notifier_ref: dict = {}
+
+    def _on_canary_pause(distance: int) -> None:
+        audit.log({"ts": time.time(), "event": "canary_drift_paused", "distance": distance})
+        n = _notifier_ref.get("n")
+        if n is not None:
+            n.send(Alert(
+                kind="warn",
+                title=f"Canary drift={distance} — monitorizare pauzată",
+                body=(
+                    "Fereastra/paleta s-a schimbat. Trimite /resume pentru a relua "
+                    "sau recalibrează."
+                ),
+            ))
+
+    canary = Canary(
+        cfg,
+        pause_flag_path=Path("logs/pause.flag"),
+        on_pause_callback=_on_canary_pause,
+    )
     detection_log = AuditLog(Path("logs/detections"))
     backends = [
         DiscordNotifier(cfg.discord.webhook_url),
@@ -521,7 +1145,6 @@ def run_live(cfg, duration_s=None, capture_stub: bool = False) -> None:
     ]
 
     def _on_drop(backend_name: str, dropped: Alert) -> None:
-        """Audit la depășire coadă — face eșecul silențios vizibil."""
         audit.log({
             "ts": time.time(),
             "event": "queue_overflow_drop",
@@ -531,8 +1154,9 @@ def run_live(cfg, duration_s=None, capture_stub: bool = False) -> None:
         })
 
     notifier = FanoutNotifier(backends, Path(cfg.dead_letter_path), on_drop=_on_drop)
+    _notifier_ref["n"] = notifier
 
-    # Verificare inițială: captură un frame, confirmă că canary se potrivește cu calibrarea.
+    # Initial frame + canary check
     first_frame = capture()
     if first_frame is None:
         print("WARN: first capture returned None — window/region missing", flush=True)
@@ -542,9 +1166,9 @@ def run_live(cfg, duration_s=None, capture_stub: bool = False) -> None:
         canary_status = f"drift={first_check.distance}/{cfg.canary.drift_threshold}"
         if first_check.drifted:
             print(f"WARN: canary drift at startup ({canary_status}). Wrong window in front?", flush=True)
-            canary.resume()  # clear the auto-pause so user can Ctrl+C and fix
+            canary.resume()
 
-    dur_note = f" dur=∞" if duration_s is None else f" dur={duration_s/3600:.2f}h"
+    dur_note = " dur=∞" if duration_s is None else f" dur={duration_s/3600:.2f}h"
     notifier.send(Alert(
         kind="heartbeat",
         title="ATM pornit",
@@ -556,106 +1180,55 @@ def run_live(cfg, duration_s=None, capture_stub: bool = False) -> None:
     audit.log({"event": "started", "config": cfg.config_version, "canary": canary_status})
 
     start = time.monotonic()
-    heartbeat_due = time.time() + cfg.heartbeat_min * 60
-    levels_extractor = None
-    last_saved_color: str | None = None
-    first_accepted = True
+    heartbeat_due = time.monotonic() + cfg.heartbeat_min * 60
     samples_dir = Path("samples")
     samples_dir.mkdir(exist_ok=True)
     fires_dir = Path("logs") / "fires"
     fires_dir.mkdir(parents=True, exist_ok=True)
-    import cv2  # type: ignore[import-untyped]
-
     try:
-        while duration_s is None or (time.monotonic() - start) < duration_s:
-            now = time.time()
-            frame = capture()
-            if frame is None:
-                audit.log({"ts": now, "event": "window_lost"})
-                time.sleep(cfg.loop_interval_s)
-                continue
-            # canary check
-            cr = canary.check(frame)
-            if canary.is_paused:
-                audit.log({"ts": now, "event": "paused", "drift": cr.distance})
-                time.sleep(cfg.loop_interval_s)
-                continue
-            # detection
-            res = detector.step(now)
-            detection_log.log({
-                "ts": now,
-                "event": "frame",
-                "window_found": res.window_found,
-                "dot_found": res.dot_found,
-                "rgb": list(res.rgb) if res.rgb is not None else None,
-                "match_name": res.match.name if res.match is not None else None,
-                "distance": round(res.match.distance, 2) if res.match is not None else None,
-                "confidence": round(res.match.confidence, 3) if res.match is not None else None,
-                "accepted": res.accepted,
-                "color": res.color,
-            })
-            if res.accepted and res.color:
-                is_first = first_accepted
-                first_accepted = False
+        import cv2  # noqa: F401  fail fast if cv2 is missing  # type: ignore[import-untyped]
+    except ImportError:
+        pass
 
-                # Per-iteration closure — binds current frame/now, gates on config.
-                def _snapshot(kind: str, label: str) -> "Path | None":
-                    if not getattr(cfg.attach_screenshots, kind, True):
-                        return None
-                    return _save_annotated_frame(
-                        frame, cfg, fires_dir, label, now, audit=audit,
-                    )
+    cmd_queue: asyncio.Queue[Command] = asyncio.Queue()
+    loop_state = _LoopState(first_accepted=True, last_saved_color=None,
+                            levels_extractor=None, fire_count=0, start=start)
 
-                tr = _handle_tick(
-                    fsm, res.color, now, notifier, audit, is_first,
-                    snapshot=_snapshot,
-                )
-                if tr is None:
-                    # pornire târzie: FSM neatins, sari peste FIRE + salvare corpus
-                    time.sleep(cfg.loop_interval_s)
-                    continue
-                # corpus: salvează frame complet la fiecare culoare nouă distinctă, pt etichetare ulterioară
-                if res.color != last_saved_color:
-                    ts_str = datetime.fromtimestamp(now).strftime("%Y%m%d_%H%M%S")
-                    sample_path = samples_dir / f"{ts_str}_{res.color}.png"
-                    try:
-                        cv2.imwrite(str(sample_path), frame)
-                    except Exception as exc:
-                        audit.log({"ts": now, "event": "sample_save_failed", "error": str(exc)})
-                    last_saved_color = res.color
-                # FIRE: adnotează frame-ul + salvează, atașează la alertă
-                if tr.trigger and not tr.locked:
-                    fire_path: "Path | None" = None
-                    if cfg.attach_screenshots.trigger:
-                        fire_path = _save_annotated_frame(
-                            frame, cfg, fires_dir, tr.trigger, now, audit=audit,
-                        )
-                    notifier.send(Alert(
-                        kind="trigger",
-                        title=f"Semnal {tr.trigger}",
-                        body=f"@ {datetime.fromtimestamp(now).isoformat(timespec='seconds')}",
-                        image_path=fire_path,
-                        direction=tr.trigger,
-                    ))
-                    levels_extractor = LevelsExtractor(cfg, tr.trigger, now)
-            # phase-B levels
-            if levels_extractor is not None:
-                lr = levels_extractor.step(frame, now)
-                if lr.status in ("complete", "timeout"):
-                    if lr.status == "complete" and lr.levels:
-                        notifier.send(Alert(
-                            kind="levels",
-                            title="Niveluri",
-                            body=(
-                                f"SL={lr.levels.sl} "
-                                f"TP1={lr.levels.tp1} "
-                                f"TP2={lr.levels.tp2}"
-                            ),
-                        ))
-                    levels_extractor = None
-            # heartbeat — include statistici per-backend ca eșecurile silențioase
-            # să apară la fiecare 30 min fără să aștepte oprirea.
-            if time.time() > heartbeat_due:
+    def _bound_save(frame: Any, label: str, now: float) -> "Path | None":
+        return _save_annotated_frame(frame, cfg, fires_dir, label, now, audit=audit)
+
+    scheduler = ScreenshotScheduler(
+        capture=capture,
+        save_fn=_bound_save,
+        notifier=notifier,
+        audit=audit,
+    )
+    poller = TelegramPoller(cfg.telegram, cmd_queue, audit)
+
+    lifecycle = LifecycleState()
+    # Seed lifecycle.last_window_state with the current status so we don't emit
+    # a spurious market_open alert on the very first tick (R2).
+    _pre_skip = _should_skip(time.time(), lifecycle, cfg, canary)
+    _maybe_log_transition(_pre_skip, lifecycle, time.time(), audit, notifier)
+
+    ctx = RunContext(
+        cfg=cfg, capture=capture, canary=canary, detector=detector, fsm=fsm,
+        notifier=notifier, audit=audit, detection_log=detection_log,
+        scheduler=scheduler, samples_dir=samples_dir, fires_dir=fires_dir,
+        cmd_queue=cmd_queue, state=loop_state,
+        levels_extractor_factory=lambda _cfg, trigger, now_ts: LevelsExtractor(_cfg, trigger, now_ts),
+        lifecycle=lifecycle,
+    )
+
+    # ------------------------------------------------------------------
+    # Nested async coroutines — heartbeat captures notifier + heartbeat_due
+    # ------------------------------------------------------------------
+
+    async def _heartbeat_loop() -> None:
+        nonlocal heartbeat_due
+        while True:
+            await asyncio.sleep(60)
+            if time.monotonic() > heartbeat_due:
                 try:
                     stats = notifier.stats()
                     audit.log({"ts": time.time(), "event": "notifier_stats", "stats": stats})
@@ -668,9 +1241,40 @@ def run_live(cfg, duration_s=None, capture_stub: bool = False) -> None:
                     notifier.send(Alert(kind="heartbeat", title="activ", body="\n".join(body_lines)))
                 except Exception:
                     notifier.send(Alert(kind="heartbeat", title="activ", body="încredere ok"))
-                heartbeat_due = time.time() + cfg.heartbeat_min * 60
-            time.sleep(cfg.loop_interval_s)
+                heartbeat_due = time.monotonic() + cfg.heartbeat_min * 60
+
+    async def _detection_loop() -> None:
+        while True:
+            if duration_s is not None and (time.monotonic() - start) >= duration_s:
+                break
+            result = await _run_tick(ctx)
+            await _handle_fsm_result(ctx, result)
+            await _drain_cmd_queue(ctx)  # UNCONDITIONAL — fix for command hang
+            await asyncio.sleep(cfg.loop_interval_s)
+
+    # Launch background tasks
+    t_scheduler = asyncio.create_task(scheduler.run(), name="scheduler")
+    t_poller = asyncio.create_task(poller.run(), name="poller")
+    t_heartbeat = asyncio.create_task(_heartbeat_loop(), name="heartbeat")
+
+    try:
+        await _detection_loop()
     finally:
+        # 7-step graceful shutdown
+        # 1. cancel scheduler
+        t_scheduler.cancel()
+        with contextlib.suppress(asyncio.CancelledError, Exception):
+            await t_scheduler
+        # 2. cancel poller
+        t_poller.cancel()
+        with contextlib.suppress(asyncio.CancelledError, Exception):
+            await t_poller
+        # 3. cancel heartbeat
+        t_heartbeat.cancel()
+        with contextlib.suppress(asyncio.CancelledError, Exception):
+            await t_heartbeat
+        # 4. drain detection — complete (we awaited _detection_loop directly)
+        # 5. send shutdown alert
         try:
             stats = notifier.stats()
             lines = [f"după {time.monotonic() - start:.0f}s"]
@@ -679,13 +1283,12 @@ def run_live(cfg, duration_s=None, capture_stub: bool = False) -> None:
                     f"{name}: sent={s['sent']} failed={s['failed']} "
                     f"dropped={s['dropped']} retries={s['retries']}"
                 )
-            notifier.send(Alert(
-                kind="heartbeat", title="ATM oprit",
-                body="\n".join(lines),
-            ))
+            notifier.send(Alert(kind="heartbeat", title="ATM oprit", body="\n".join(lines)))
         except Exception:
             pass
+        # 6. notifier.stop() — flush + join FanoutNotifier threads
         notifier.stop()
+        # 7. audit.close()
         audit.close()
         detection_log.close()
 
diff --git a/src/atm/notifier/__init__.py b/src/atm/notifier/__init__.py
index c6afeb5..2a61f79 100644
--- a/src/atm/notifier/__init__.py
+++ b/src/atm/notifier/__init__.py
@@ -5,11 +5,13 @@ from typing import Protocol
 
 @dataclass
 class Alert:
-    kind: str          # "trigger" | "heartbeat" | "levels" | "warn" | "arm" | "prime" | "late_start"
+    # flat union: "trigger"|"heartbeat"|"levels"|"warn"|"arm"|"prime"|"late_start"|"screenshot"|"status"
+    kind: str
     title: str
     body: str
     image_path: Path | None = None  # annotated screenshot
     direction: str | None = None    # "BUY"/"SELL" when kind=trigger
+    silent: bool = False            # disable_notification for Telegram; ignored by Discord
 
 
 class Notifier(Protocol):
diff --git a/src/atm/notifier/telegram.py b/src/atm/notifier/telegram.py
index ad5db2f..0e5b366 100644
--- a/src/atm/notifier/telegram.py
+++ b/src/atm/notifier/telegram.py
@@ -33,6 +33,7 @@ class TelegramNotifier:
                         "chat_id": self._chat_id,
                         "caption": text,
                         "parse_mode": "HTML",
+                        "disable_notification": str(alert.silent).lower(),
                     },
                     files={"photo": fh},
                     timeout=10,
@@ -44,6 +45,7 @@ class TelegramNotifier:
                     "chat_id": self._chat_id,
                     "text": text,
                     "parse_mode": "HTML",
+                    "disable_notification": alert.silent,
                 },
                 timeout=10,
             )
diff --git a/src/atm/scheduler.py b/src/atm/scheduler.py
new file mode 100644
index 0000000..2757e6f
--- /dev/null
+++ b/src/atm/scheduler.py
@@ -0,0 +1,118 @@
+"""ScreenshotScheduler — periodic capture + annotate + send.
+
+Runs as an asyncio task. capture() and cv2 work execute in asyncio.to_thread
+to avoid blocking the event loop. Decision 13: scheduler calls capture()
+directly, NOT via Detector.
+"""
+from __future__ import annotations
+
+import asyncio
+import logging
+import time
+from pathlib import Path
+from typing import Callable
+
+from .notifier import Alert
+
+logger = logging.getLogger(__name__)
+
+
+class ScreenshotScheduler:
+    """Periodic screenshot sender.
+
+    Constructor params are explicit (decision 11 outside-voice finding).
+    """
+
+    def __init__(
+        self,
+        capture: Callable,               # () -> ndarray | None
+        save_fn: Callable,               # (frame, label, now) -> Path | None
+        notifier,                        # _NotifierLike
+        audit,                           # _AuditLike
+        interval_s: int | None = None,
+    ) -> None:
+        self._capture = capture
+        self._save_fn = save_fn
+        self._notifier = notifier
+        self._audit = audit
+        self._interval_s = interval_s
+        self._is_running = False
+        self._next_due: float | None = None  # monotonic
+
+    # ------------------------------------------------------------------
+    # Public state
+    # ------------------------------------------------------------------
+
+    @property
+    def is_running(self) -> bool:
+        return self._is_running
+
+    @property
+    def interval_s(self) -> int | None:
+        return self._interval_s
+
+    @property
+    def next_due(self) -> float | None:
+        return self._next_due
+
+    # ------------------------------------------------------------------
+    # Control (called from async event loop)
+    # ------------------------------------------------------------------
+
+    def start(self, interval_s: int) -> None:
+        self._interval_s = interval_s
+        self._is_running = True
+        self._next_due = time.monotonic() + interval_s
+
+    def stop(self) -> None:
+        self._is_running = False
+        self._next_due = None
+
+    # ------------------------------------------------------------------
+    # Task body
+    # ------------------------------------------------------------------
+
+    async def run(self) -> None:
+        """Runs until cancelled."""
+        while True:
+            await asyncio.sleep(1)
+            if not self._is_running or self._next_due is None:
+                continue
+            if time.monotonic() >= self._next_due:
+                await self._take_screenshot()
+                if self._is_running and self._interval_s is not None:
+                    self._next_due = time.monotonic() + self._interval_s
+
+    async def _take_screenshot(self) -> None:
+        now = time.time()
+        try:
+            frame = await asyncio.to_thread(self._capture)
+        except Exception as exc:
+            logger.warning("ScreenshotScheduler capture failed: %s", exc)
+            self._audit.log({"ts": now, "event": "screenshot_sent", "status": "capture_failed", "error": str(exc)})
+            self._notifier.send(Alert(
+                kind="warn",
+                title="Captură eșuată — verificați fereastra TradeStation",
+                body="",
+                silent=True,
+            ))
+            return
+
+        if frame is None:
+            self._notifier.send(Alert(
+                kind="warn",
+                title="Captură eșuată — verificați fereastra TradeStation",
+                body="",
+                silent=True,
+            ))
+            return
+
+        path = await asyncio.to_thread(self._save_fn, frame, "poll", now)
+        self._audit.log({"ts": now, "event": "screenshot_sent", "path": str(path) if path else None})
+        self._notifier.send(Alert(
+            kind="screenshot",
+            title="Screenshot periodic",
+            body="",
+            image_path=path,
+            silent=True,
+        ))
diff --git a/src/atm/state_machine.py b/src/atm/state_machine.py
index f4e92d2..88cc2f8 100644
--- a/src/atm/state_machine.py
+++ b/src/atm/state_machine.py
@@ -232,3 +232,20 @@ class StateMachine:
         if last is None:
             return False
         return (ts - last) < self._lockout_s
+
+    # ------------------------------------------------------------------
+    # Public lockout API — used by fire_on_phase_skip handler outside the
+    # FSM. Mirrors _is_locked / _last_fire without leaking private attrs.
+    # ------------------------------------------------------------------
+
+    def is_locked(self, direction: str, ts: float) -> bool:
+        """True if a FIRE in `direction` at ts would be within the lockout window."""
+        return self._is_locked(direction, ts)
+
+    def record_fire(self, direction: str, ts: float) -> None:
+        """Mark a FIRE for `direction` at ts, starting the lockout timer.
+
+        Used by backstop handlers (e.g. fire_on_phase_skip) that emit a
+        fire-equivalent alert without going through the natural FSM path.
+        """
+        self._last_fire[direction] = ts
diff --git a/src/atm/validate.py b/src/atm/validate.py
new file mode 100644
index 0000000..2b89ff8
--- /dev/null
+++ b/src/atm/validate.py
@@ -0,0 +1,229 @@
+"""Offline calibration validation: run Detector on labeled frames, report PASS/FAIL.
+
+Used by the `atm validate-calibration` subcommand. Reports per-sample detection
+results against expected labels, and for failures, computes RGB distance to
+each color threshold and emits tuning suggestions.
+
+Reuses `Detector.step(frame)` - does NOT reimplement color classification.
+"""
+from __future__ import annotations
+
+import json
+import math
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+from .config import Config
+
+
+@dataclass
+class SampleRecord:
+    path: str
+    expected: str
+    detected: str | None
+    confidence: float
+    rgb: tuple[int, int, int] | None
+    top3: list[tuple[str, float]]  # [(name, score), ...] ranked by RGB distance
+    passed: bool
+    note: str = ""
+    error: str | None = None  # non-None if frame load failed / schema bad
+
+
+@dataclass
+class ValidationReport:
+    records: list[SampleRecord] = field(default_factory=list)
+    config_name: str = ""
+
+    @property
+    def total(self) -> int:
+        return len(self.records)
+
+    @property
+    def passed(self) -> int:
+        return sum(1 for r in self.records if r.passed)
+
+    @property
+    def failed(self) -> int:
+        return self.total - self.passed
+
+    @property
+    def all_pass(self) -> bool:
+        return self.total > 0 and self.failed == 0
+
+    def render(self) -> str:
+        lines: list[str] = []
+        hdr = f"Testing {self.total} frames"
+        if self.config_name:
+            hdr += f" against config {self.config_name}"
+        hdr += "..."
+        lines.append(hdr)
+        lines.append("")
+
+        for r in self.records:
+            name = Path(r.path).name or r.path
+            if r.error:
+                lines.append(f"  [FAIL] {name}")
+                lines.append(f"         error: {r.error}")
+                continue
+            tag = "PASS" if r.passed else "FAIL"
+            rgb_str = f"RGB {r.rgb}" if r.rgb is not None else "RGB n/a"
+            detected = r.detected if r.detected is not None else "none"
+            lines.append(f"  [{tag}] {name}")
+            lines.append(
+                f"         expected={r.expected}  detected={detected}  "
+                f"(conf {r.confidence:.2f}, {rgb_str})"
+            )
+            if not r.passed and r.top3:
+                top3_str = " ".join(f"{n}({c:.2f})" for n, c in r.top3)
+                lines.append(f"         Top 3 candidates: {top3_str}")
+
+        lines.append("")
+        pct = (self.passed / self.total * 100.0) if self.total else 0.0
+        lines.append(f"SUMMARY: {self.passed}/{self.total} PASS ({pct:.0f}%)")
+
+        fails = [r for r in self.records if not r.passed]
+        if fails:
+            lines.append("FAILED:")
+            for r in fails:
+                name = Path(r.path).name or r.path
+                if r.error:
+                    lines.append(f"  - {name}: {r.error}")
+                    continue
+                detected = r.detected if r.detected is not None else "none"
+                lines.append(
+                    f"  - {name}: expected {r.expected}, got {detected}"
+                )
+
+            sug_lines = [
+                r._suggestion  # type: ignore[attr-defined]
+                for r in fails
+                if getattr(r, "_suggestion", "")
+            ]
+            if sug_lines:
+                lines.append("")
+                lines.append("SUGGESTIONS:")
+                for s in sug_lines:
+                    lines.append(f"  - {s}")
+
+        return "\n".join(lines)
+
+    def __str__(self) -> str:
+        return self.render()
+
+
+class ValidationError(Exception):
+    """Raised for missing label files or invalid schema."""
+
+
+def _rgb_distance(a: tuple[int, int, int], b: tuple[int, int, int]) -> float:
+    return math.sqrt(sum((a[i] - b[i]) ** 2 for i in range(3)))
+
+
+def _load_labels(label_file: Path) -> list[dict[str, Any]]:
+    if not label_file.exists():
+        raise ValidationError(f"label file not found: {label_file}")
+    try:
+        data = json.loads(label_file.read_text(encoding="utf-8"))
+    except json.JSONDecodeError as exc:
+        raise ValidationError(f"invalid JSON in {label_file}: {exc}") from exc
+    if not isinstance(data, list):
+        raise ValidationError(
+            f"label file must be a JSON array; got {type(data).__name__}"
+        )
+    return data
+
+
+def validate_calibration(
+    label_file: Path,
+    cfg: Config,
+    config_name: str = "",
+) -> ValidationReport:
+    """Run Detector on each labeled frame; return a ValidationReport.
+
+    Reuses `Detector.step(frame)`. Loads frames via cv2.imread.
+    Raises ValidationError if the label file is missing or malformed.
+    """
+    import cv2  # local import keeps module import cheap
+    from .detector import Detector
+
+    entries = _load_labels(label_file)
+    report = ValidationReport(config_name=config_name)
+
+    palette = {
+        name: spec.rgb
+        for name, spec in cfg.colors.items()
+        if name != "background"
+    }
+
+    detector = Detector(cfg=cfg, capture=lambda: None)
+
+    for entry in entries:
+        path = str(entry.get("path", ""))
+        expected = str(entry.get("expected", ""))
+        note = str(entry.get("note", ""))
+
+        if not path or not expected:
+            rec = SampleRecord(
+                path=path, expected=expected, detected=None, confidence=0.0,
+                rgb=None, top3=[], passed=False, note=note,
+                error="missing 'path' or 'expected' field",
+            )
+            rec._suggestion = ""  # type: ignore[attr-defined]
+            report.records.append(rec)
+            continue
+
+        frame = cv2.imread(path)
+        if frame is None:
+            rec = SampleRecord(
+                path=path, expected=expected, detected=None, confidence=0.0,
+                rgb=None, top3=[], passed=False, note=note,
+                error=f"cv2.imread failed for {path}",
+            )
+            rec._suggestion = ""  # type: ignore[attr-defined]
+            report.records.append(rec)
+            continue
+
+        result = detector.step(ts=0.0, frame=frame)
+
+        match = result.match
+        if match is None:
+            detected: str | None = None
+            confidence = 0.0
+        else:
+            detected = match.name if match.name != "UNKNOWN" else None
+            confidence = match.confidence
+
+        rgb = result.rgb
+
+        # Top 3 candidates: rank palette entries by RGB distance to observed.
+        top3: list[tuple[str, float]] = []
+        if rgb is not None:
+            scored: list[tuple[str, float]] = []
+            for name, ref in palette.items():
+                scored.append((name, _rgb_distance(rgb, ref)))
+            scored.sort(key=lambda t: t[1])
+            top3 = [(n, 1.0 / (1.0 + d / 20.0)) for n, d in scored[:3]]
+
+        passed = detected == expected
+
+        rec = SampleRecord(
+            path=path, expected=expected, detected=detected,
+            confidence=confidence, rgb=rgb, top3=top3, passed=passed, note=note,
+        )
+
+        if not passed and rgb is not None and expected in palette:
+            ref = palette[expected]
+            tol = cfg.colors[expected].tolerance
+            dist = _rgb_distance(rgb, ref)
+            rec._suggestion = (  # type: ignore[attr-defined]
+                f"{expected} praguri curente: RGB{ref} +/- {tol:.0f}. "
+                f"Pixelul observat {rgb} e la distanta {dist:.1f} "
+                f"-> recalibreaza cu acest sample."
+            )
+        else:
+            rec._suggestion = ""  # type: ignore[attr-defined]
+
+        report.records.append(rec)
+
+    return report
diff --git a/tests/test_canary.py b/tests/test_canary.py
index feb63a8..4dcce37 100644
--- a/tests/test_canary.py
+++ b/tests/test_canary.py
@@ -140,6 +140,52 @@ def test_pause_file_written(tmp_path: Path) -> None:
     assert flag.exists()
 
 
+def test_canary_pause_callback_fires_once() -> None:
+    """Single-shot: callback invoked exactly once per not_paused→paused edge."""
+    cfg = _cfg_with_baseline(BASELINE_FRAME)
+    calls: list[int] = []
+
+    canary = Canary(cfg, on_pause_callback=lambda d: calls.append(d))
+
+    canary.check(DRIFTED_FRAME)   # transition → callback fires
+    canary.check(DRIFTED_FRAME)   # still paused → no new callback
+    canary.check(BASELINE_FRAME)  # clean but still paused → no new callback
+
+    assert len(calls) == 1
+    assert calls[0] > 0  # distance should be positive
+
+
+def test_canary_resume_allows_new_pause_notification() -> None:
+    """After resume, a fresh drift must re-fire the callback."""
+    cfg = _cfg_with_baseline(BASELINE_FRAME)
+    calls: list[int] = []
+
+    canary = Canary(cfg, on_pause_callback=lambda d: calls.append(d))
+
+    canary.check(DRIFTED_FRAME)
+    assert len(calls) == 1
+
+    canary.resume()
+    canary.check(DRIFTED_FRAME)  # new pause transition
+
+    assert len(calls) == 2
+
+
+def test_canary_pause_callback_exception_does_not_crash_check() -> None:
+    """A failing callback must not break canary.check (detection cycle safety)."""
+    cfg = _cfg_with_baseline(BASELINE_FRAME)
+
+    def _boom(_d: int) -> None:
+        raise RuntimeError("notifier down")
+
+    canary = Canary(cfg, on_pause_callback=_boom)
+
+    # Must not raise — exception is swallowed + logged.
+    result = canary.check(DRIFTED_FRAME)
+    assert result.paused is True
+    assert canary.is_paused is True
+
+
 def test_resume_deletes_pause_file(tmp_path: Path) -> None:
     """resume() deletes the pause flag file."""
     flag = tmp_path / "paused.flag"
diff --git a/tests/test_commands.py b/tests/test_commands.py
new file mode 100644
index 0000000..7a9bdf6
--- /dev/null
+++ b/tests/test_commands.py
@@ -0,0 +1,45 @@
+"""Tests for atm.commands — /pause /resume parsing (Commit 5)."""
+from __future__ import annotations
+
+from unittest.mock import MagicMock
+
+from atm.commands import Command, TelegramPoller
+
+
+def _make_poller() -> TelegramPoller:
+    cfg = MagicMock()
+    cfg.bot_token = "tok"
+    cfg.chat_id = "123"
+    cfg.allowed_chat_ids = ("123",)
+    cfg.poll_timeout_s = 1
+    return TelegramPoller(cfg, MagicMock(), MagicMock())
+
+
+def test_parse_pause():
+    p = _make_poller()
+    assert p._parse_command("pause") == Command(action="pause")
+    assert p._parse_command("/pause") == Command(action="pause")
+
+
+def test_parse_resume_plain():
+    p = _make_poller()
+    assert p._parse_command("resume") == Command(action="resume")
+    assert p._parse_command("/resume") == Command(action="resume")
+
+
+def test_parse_resume_force():
+    p = _make_poller()
+    # "resume force" → value=1 signals force-resume of canary drift
+    cmd = p._parse_command("resume force")
+    assert cmd is not None
+    assert cmd.action == "resume"
+    assert cmd.value == 1
+
+
+def test_parse_existing_commands_still_work():
+    """Regression: adding pause/resume must not break stop/status/ss/interval."""
+    p = _make_poller()
+    assert p._parse_command("stop") == Command(action="stop")
+    assert p._parse_command("status") == Command(action="status")
+    assert p._parse_command("ss") == Command(action="ss")
+    assert p._parse_command("3") == Command(action="set_interval", value=180)
diff --git a/tests/test_config.py b/tests/test_config.py
index f864366..6825ced 100644
--- a/tests/test_config.py
+++ b/tests/test_config.py
@@ -97,3 +97,59 @@ def test_attach_screenshots_unknown_keys_ignored() -> None:
     }))
     assert cfg.attach_screenshots.arm is False
     # Should not raise even with unknown key
+
+
+# ---------------------------------------------------------------------------
+# Commit 3: AlertBehaviorCfg (fire_on_phase_skip)
+# ---------------------------------------------------------------------------
+
+def test_alerts_default_fire_on_phase_skip_true() -> None:
+    cfg = Config._from_dict(_with_opts({}))
+    assert cfg.alerts.fire_on_phase_skip is True
+
+
+def test_alerts_fire_on_phase_skip_can_be_disabled() -> None:
+    cfg = Config._from_dict(_with_opts({"alerts": {"fire_on_phase_skip": False}}))
+    assert cfg.alerts.fire_on_phase_skip is False
+
+
+# ---------------------------------------------------------------------------
+# Commit 4: OperatingHoursCfg parsing + tz cache
+# ---------------------------------------------------------------------------
+
+def test_operating_hours_default_disabled() -> None:
+    cfg = Config._from_dict(_with_opts({}))
+    assert cfg.operating_hours.enabled is False
+    assert cfg.operating_hours.timezone == "America/New_York"
+    assert cfg.operating_hours._tz_cache is None
+
+
+def test_operating_hours_enabled_caches_tz() -> None:
+    cfg = Config._from_dict(_with_opts({
+        "operating_hours": {
+            "enabled": True,
+            "timezone": "America/New_York",
+            "weekdays": ["MON", "TUE", "WED", "THU", "FRI"],
+            "start_hhmm": "09:30",
+            "stop_hhmm": "16:00",
+        }
+    }))
+    assert cfg.operating_hours.enabled is True
+    assert cfg.operating_hours._tz_cache is not None
+    assert str(cfg.operating_hours._tz_cache) == "America/New_York"
+
+
+def test_operating_hours_invalid_tz_raises_valueerror() -> None:
+    import pytest
+    with pytest.raises(ValueError, match="operating_hours.timezone"):
+        Config._from_dict(_with_opts({
+            "operating_hours": {"enabled": True, "timezone": "Not/A_Zone"},
+        }))
+
+
+def test_operating_hours_invalid_weekday_raises_valueerror() -> None:
+    import pytest
+    with pytest.raises(ValueError, match="weekdays"):
+        Config._from_dict(_with_opts({
+            "operating_hours": {"enabled": True, "weekdays": ["XYZ"]},
+        }))
diff --git a/tests/test_handle_tick.py b/tests/test_handle_tick.py
index 958ce70..943a5c9 100644
--- a/tests/test_handle_tick.py
+++ b/tests/test_handle_tick.py
@@ -10,6 +10,8 @@ Covers the six cases from the arm+prime notification plan:
 """
 from __future__ import annotations
 
+from types import SimpleNamespace
+
 from atm.main import _handle_tick
 from atm.notifier import Alert
 from atm.state_machine import State, StateMachine
@@ -486,3 +488,82 @@ def test_save_annotated_frame_succeeds(tmp_path, monkeypatch):
     assert "BUY" in result.name
     assert len(written) == 1
     assert not any(e.get("event") == "snapshot_fail" for e in audit.events)
+
+
+# ---------------------------------------------------------------------------
+# Commit 3: fire_on_phase_skip backstop
+# ---------------------------------------------------------------------------
+
+def _cfg_with_flag(enabled: bool):
+    return SimpleNamespace(alerts=SimpleNamespace(fire_on_phase_skip=enabled))
+
+
+def test_phase_skip_fire_when_flag_on():
+    """ARMED_SELL → light_red directly with flag=True → phase_skip_fire alert."""
+    fsm = StateMachine(lockout_s=240)
+    notif = FakeNotifier()
+    audit = FakeAudit()
+
+    # Arm SELL (yellow from IDLE)
+    _handle_tick(fsm, "yellow", 1.0, notif, audit, first_accepted=False,
+                 cfg=_cfg_with_flag(True))
+    assert fsm.state == State.ARMED_SELL
+    notif.alerts.clear()
+
+    # ARMED_SELL → light_red (skips dark_red) → phase_skip_fire
+    tr = _handle_tick(fsm, "light_red", 2.0, notif, audit, first_accepted=False,
+                      cfg=_cfg_with_flag(True))
+    assert tr is not None and tr.reason == "phase_skip"
+
+    ps_alerts = [a for a in notif.alerts if a.kind == "phase_skip_fire"]
+    assert len(ps_alerts) == 1
+    assert ps_alerts[0].direction == "SELL"
+    assert "SELL" in ps_alerts[0].title
+
+
+def test_phase_skip_no_fire_when_flag_off():
+    """Same scenario, flag=False → no phase_skip_fire emitted."""
+    fsm = StateMachine(lockout_s=240)
+    notif = FakeNotifier()
+    audit = FakeAudit()
+
+    _handle_tick(fsm, "yellow", 1.0, notif, audit, first_accepted=False,
+                 cfg=_cfg_with_flag(False))
+    notif.alerts.clear()
+
+    _handle_tick(fsm, "light_red", 2.0, notif, audit, first_accepted=False,
+                 cfg=_cfg_with_flag(False))
+
+    ps_alerts = [a for a in notif.alerts if a.kind == "phase_skip_fire"]
+    assert ps_alerts == []
+
+
+def test_phase_skip_lockout_suppresses_spam():
+    """Two phase_skip events within lockout_s → only the first emits an alert."""
+    fsm = StateMachine(lockout_s=240)
+    notif = FakeNotifier()
+    audit = FakeAudit()
+    cfg = _cfg_with_flag(True)
+
+    # First cycle
+    _handle_tick(fsm, "yellow", 1.0, notif, audit, first_accepted=False, cfg=cfg)
+    _handle_tick(fsm, "light_red", 2.0, notif, audit, first_accepted=False, cfg=cfg)
+    # Second arm + phase_skip well within 240s
+    _handle_tick(fsm, "yellow", 60.0, notif, audit, first_accepted=False, cfg=cfg)
+    _handle_tick(fsm, "light_red", 61.0, notif, audit, first_accepted=False, cfg=cfg)
+
+    ps_alerts = [a for a in notif.alerts if a.kind == "phase_skip_fire"]
+    assert len(ps_alerts) == 1, (
+        f"expected 1 phase_skip_fire (lockout), got {len(ps_alerts)}"
+    )
+
+
+def test_state_machine_is_locked_and_record_fire_public_api():
+    """Public lockout helpers mirror the private _is_locked / _last_fire behavior."""
+    fsm = StateMachine(lockout_s=100)
+    assert fsm.is_locked("BUY", 0.0) is False
+
+    fsm.record_fire("BUY", 10.0)
+    assert fsm.is_locked("BUY", 50.0) is True   # within 100s
+    assert fsm.is_locked("BUY", 150.0) is False  # past lockout
+    assert fsm.is_locked("SELL", 50.0) is False  # other direction unaffected
diff --git a/tests/test_main.py b/tests/test_main.py
index ee8e18e..0901643 100644
--- a/tests/test_main.py
+++ b/tests/test_main.py
@@ -1,6 +1,7 @@
 """Tests for atm.main unified CLI."""
 from __future__ import annotations
 
+import asyncio
 import os
 import subprocess
 import sys
@@ -186,7 +187,7 @@ def test_run_live_catchup_sell_from_gray_then_dark_red(monkeypatch, tmp_path):
         ]
         def __init__(self, *a, **kw):
             self._i = 0
-        def step(self, ts):
+        def step(self, ts, frame=None):
             if self._i >= len(self._script):
                 raise _StopLoop
             color, accepted = self._script[self._i]
@@ -228,6 +229,17 @@ def test_run_live_catchup_sell_from_gray_then_dark_red(monkeypatch, tmp_path):
         def step(self, *a, **kw):
             return types.SimpleNamespace(status="pending", levels=None)
 
+    class _StubPoller:
+        def __init__(self, *a, **kw): pass
+        async def run(self): await asyncio.sleep(9999)
+
+    class _StubScheduler:
+        def __init__(self, *a, **kw):
+            self.is_running = False
+        def start(self, interval_s): self.is_running = True
+        def stop(self): self.is_running = False
+        async def run(self): await asyncio.sleep(9999)
+
     monkeypatch.setattr("atm.detector.Detector", ScriptedDetector)
     monkeypatch.setattr("atm.canary.Canary", FakeCanary)
     monkeypatch.setattr("atm.notifier.fanout.FanoutNotifier", FakeFanout)
@@ -237,6 +249,8 @@ def test_run_live_catchup_sell_from_gray_then_dark_red(monkeypatch, tmp_path):
     monkeypatch.setattr("atm.levels.LevelsExtractor", _Stub)
     monkeypatch.setattr("atm.main._build_capture", fake_build_capture)
     monkeypatch.setattr("atm.main.time.sleep", lambda s: None)
+    monkeypatch.setattr("atm.commands.TelegramPoller", _StubPoller)
+    monkeypatch.setattr("atm.scheduler.ScreenshotScheduler", _StubScheduler)
 
     with pytest.raises(_StopLoop):
         _main.run_live(cfg, duration_s=None)
@@ -255,3 +269,713 @@ def test_run_live_catchup_sell_from_gray_then_dark_red(monkeypatch, tmp_path):
 
     assert len(trigger) == 1
     assert trigger[0].direction == "SELL"
+
+
+# ---------------------------------------------------------------------------
+# MUST-HAVE: async lifecycle integration test
+# IDLE → ARMED → PRIMED (auto-poll scheduler starts) → FIRE (scheduler stops)
+# Tests: scheduler starts on prime, stops on fire, fire alert sent.
+# ---------------------------------------------------------------------------
+
+@pytest.mark.asyncio
+async def test_lifecycle_idle_armed_primed_autopoll_fire_stop(monkeypatch, tmp_path):
+    import numpy as np
+    import atm.main as _main
+    from atm.detector import DetectionResult
+
+    captured_alerts: list = []
+    scheduler_events: list[str] = []
+
+    class FakeFanout:
+        def __init__(self, *a, **kw): pass
+        def send(self, alert): captured_alerts.append(alert)
+        def stop(self): pass
+        def stats(self): return {}
+
+    class FakeCanaryResult:
+        distance = 0
+        drifted = False
+        paused = False
+
+    class FakeCanary:
+        def __init__(self, *a, **kw): self.is_paused = False
+        def check(self, frame): return FakeCanaryResult()
+        def resume(self): pass
+
+    # Scheduler tracks start/stop calls
+    class FakeScheduler:
+        def __init__(self, *a, **kw):
+            self.is_running = False
+            self.interval_s = None
+        def start(self, interval_s):
+            self.is_running = True
+            self.interval_s = interval_s
+            scheduler_events.append(f"start:{interval_s}")
+        def stop(self):
+            self.is_running = False
+            scheduler_events.append("stop")
+        async def run(self):
+            await asyncio.sleep(9999)
+
+    class FakePoller:
+        def __init__(self, *a, **kw): pass
+        async def run(self): await asyncio.sleep(9999)
+
+    class _StopLoop(Exception): pass
+
+    class ScriptedDetector:
+        # turquoise→ARM, dark_green→PRIME, light_green→FIRE
+        _script = [
+            ("turquoise", True),
+            ("dark_green", True),
+            ("light_green", True),
+        ]
+        def __init__(self, *a, **kw): self._i = 0
+        def step(self, ts, frame=None):
+            if self._i >= len(self._script):
+                raise _StopLoop
+            color, accepted = self._script[self._i]
+            self._i += 1
+            return DetectionResult(ts=ts, window_found=True, dot_found=True,
+                                   rgb=(1, 1, 1), match=None, accepted=accepted, color=color)
+        @property
+        def rolling(self): return []
+
+    def fake_build_capture(cfg, capture_stub=False):
+        return lambda: np.zeros((50, 50, 3), dtype=np.uint8)
+
+    cfg = MagicMock()
+    cfg.lockout_s = 60
+    cfg.heartbeat_min = 999
+    cfg.loop_interval_s = 0
+    cfg.config_version = "test"
+    cfg.dead_letter_path = str(tmp_path / "dl.jsonl")
+    cfg.canary.drift_threshold = 10
+    cfg.dot_roi.x = 0; cfg.dot_roi.y = 0; cfg.dot_roi.w = 10; cfg.dot_roi.h = 10
+    cfg.chart_window_region = None
+    cfg.telegram.auto_poll_interval_s = 180
+    cfg.telegram.bot_token = "tok"
+    cfg.telegram.chat_id = "123"
+    cfg.telegram.allowed_chat_ids = ("123",)
+
+    fake_sched = FakeScheduler()
+
+    monkeypatch.chdir(tmp_path)
+
+    class _Stub:
+        def __init__(self, *a, **kw): pass
+        def log(self, *a, **kw): pass
+        def close(self, *a, **kw): pass
+        def step(self, *a, **kw): return types.SimpleNamespace(status="pending", levels=None)
+
+    monkeypatch.setattr("atm.detector.Detector", ScriptedDetector)
+    monkeypatch.setattr("atm.canary.Canary", FakeCanary)
+    monkeypatch.setattr("atm.notifier.fanout.FanoutNotifier", FakeFanout)
+    monkeypatch.setattr("atm.notifier.discord.DiscordNotifier", _Stub)
+    monkeypatch.setattr("atm.notifier.telegram.TelegramNotifier", _Stub)
+    monkeypatch.setattr("atm.audit.AuditLog", _Stub)
+    monkeypatch.setattr("atm.levels.LevelsExtractor", _Stub)
+    monkeypatch.setattr("atm.main._build_capture", fake_build_capture)
+    monkeypatch.setattr("atm.commands.TelegramPoller", FakePoller)
+    monkeypatch.setattr("atm.scheduler.ScreenshotScheduler", lambda *a, **kw: fake_sched)
+
+    with pytest.raises(_StopLoop):
+        await _main.run_live_async(cfg, duration_s=None)
+
+    arm_alerts = [a for a in captured_alerts if a.kind == "arm"]
+    prime_alerts = [a for a in captured_alerts if a.kind == "prime"]
+    trigger_alerts = [a for a in captured_alerts if a.kind == "trigger"]
+
+    assert len(arm_alerts) == 1, f"expected 1 arm, got {[a.title for a in captured_alerts]}"
+    assert arm_alerts[0].direction == "BUY"
+
+    assert len(prime_alerts) == 1
+    assert prime_alerts[0].direction == "BUY"
+
+    assert len(trigger_alerts) == 1
+    assert trigger_alerts[0].direction == "BUY"
+
+    # Scheduler must have started (on PRIME) and stopped (on FIRE)
+    assert "start:180" in scheduler_events, f"scheduler not started: {scheduler_events}"
+    assert "stop" in scheduler_events, f"scheduler not stopped: {scheduler_events}"
+    start_idx = scheduler_events.index("start:180")
+    stop_idx = scheduler_events.index("stop")
+    assert start_idx < stop_idx, "scheduler started after it stopped"
+
+
+# ---------------------------------------------------------------------------
+# Commit 1 regression tests: _drain_cmd_queue MUST run unconditionally,
+# even when canary is paused or when detection is otherwise skipped.
+# Prior bug: `continue` past the drain loop caused commands to pile up.
+# ---------------------------------------------------------------------------
+
+def _make_ctx_for_drain(cmd_queue, dispatched: list):
+    """Build a minimal RunContext where _dispatch_command just records calls."""
+    import atm.main as _main
+
+    class _FakeAudit:
+        def __init__(self): self.events = []
+        def log(self, e): self.events.append(e)
+
+    class _FakeNotifier:
+        def __init__(self): self.alerts = []
+        def send(self, a): self.alerts.append(a)
+
+    class _FakeCanary:
+        def __init__(self, paused=True):
+            self.is_paused = paused
+
+    class _FakeScheduler:
+        is_running = False
+        interval_s = None
+        def start(self, s): pass
+        def stop(self): pass
+
+    state = _main._LoopState(start=0.0)
+    ctx = _main.RunContext(
+        cfg=MagicMock(),
+        capture=lambda: None,
+        canary=_FakeCanary(paused=True),
+        detector=MagicMock(),
+        fsm=MagicMock(),
+        notifier=_FakeNotifier(),
+        audit=_FakeAudit(),
+        detection_log=_FakeAudit(),
+        scheduler=_FakeScheduler(),
+        samples_dir=Path("."),
+        fires_dir=Path("."),
+        cmd_queue=cmd_queue,
+        state=state,
+        levels_extractor_factory=lambda *a, **kw: None,
+    )
+    return ctx
+
+
+@pytest.mark.asyncio
+async def test_drain_works_when_canary_paused(monkeypatch):
+    """Regression: when canary.is_paused, _drain_cmd_queue still dispatches.
+
+    Prior bug: detection loop `continue`'d past the drain block whenever the
+    tick returned res=None (canary paused). Commands accumulated forever.
+    """
+    import atm.main as _main
+    from atm.commands import Command
+
+    q: asyncio.Queue = asyncio.Queue()
+    await q.put(Command(action="status"))
+    await q.put(Command(action="ss"))
+
+    dispatched: list = []
+
+    async def _fake_dispatch(ctx, cmd):
+        dispatched.append(cmd.action)
+
+    monkeypatch.setattr(_main, "_dispatch_command", _fake_dispatch)
+
+    ctx = _make_ctx_for_drain(q, dispatched)
+
+    await _main._drain_cmd_queue(ctx)
+
+    assert dispatched == ["status", "ss"]
+    assert q.empty()
+
+
+@pytest.mark.asyncio
+async def test_drain_works_when_out_of_window(monkeypatch):
+    """Drain must still fire when the tick skipped (e.g. out of operating hours).
+
+    The refactored loop runs _drain_cmd_queue unconditionally after every tick,
+    regardless of `_TickSyncResult` content.
+    """
+    import atm.main as _main
+    from atm.commands import Command
+
+    q: asyncio.Queue = asyncio.Queue()
+    await q.put(Command(action="stop"))
+
+    dispatched: list = []
+
+    async def _fake_dispatch(ctx, cmd):
+        dispatched.append(cmd.action)
+
+    monkeypatch.setattr(_main, "_dispatch_command", _fake_dispatch)
+
+    ctx = _make_ctx_for_drain(q, dispatched)
+    # Simulate out-of-window tick (empty _TickSyncResult, no res)
+    await _main._handle_fsm_result(ctx, _main._TickSyncResult())
+    await _main._drain_cmd_queue(ctx)
+
+    assert dispatched == ["stop"]
+
+
+@pytest.mark.asyncio
+async def test_drain_isolates_dispatch_exceptions(monkeypatch):
+    """If one command raises, remaining commands still drain + warn alert sent."""
+    import atm.main as _main
+    from atm.commands import Command
+
+    q: asyncio.Queue = asyncio.Queue()
+    await q.put(Command(action="status"))
+    await q.put(Command(action="ss"))
+
+    attempts: list = []
+
+    async def _fake_dispatch(ctx, cmd):
+        attempts.append(cmd.action)
+        if cmd.action == "status":
+            raise RuntimeError("boom")
+
+    monkeypatch.setattr(_main, "_dispatch_command", _fake_dispatch)
+
+    ctx = _make_ctx_for_drain(q, attempts)
+    await _main._drain_cmd_queue(ctx)
+
+    assert attempts == ["status", "ss"]
+    # warn alert for the failed command
+    warn_titles = [a.title for a in ctx.notifier.alerts if a.kind == "warn"]
+    assert any("status" in t for t in warn_titles)
+    # command_error audit event
+    errs = [e for e in ctx.audit.events if e.get("event") == "command_error"]
+    assert len(errs) == 1 and errs[0]["action"] == "status"
+
+
+# ---------------------------------------------------------------------------
+# Commit 4: operating hours + LifecycleState transitions
+# ---------------------------------------------------------------------------
+
+from zoneinfo import ZoneInfo as _ZI  # noqa: E402
+import datetime as _dt  # noqa: E402
+
+
+def _oh_cfg(enabled=True, weekdays=("MON", "TUE", "WED", "THU", "FRI"),
+            start="09:30", stop="16:00", tz="America/New_York"):
+    """Build a lightweight cfg-like object with operating_hours populated."""
+    oh = types.SimpleNamespace(
+        enabled=enabled,
+        timezone=tz,
+        weekdays=weekdays,
+        start_hhmm=start,
+        stop_hhmm=stop,
+        _tz_cache=_ZI(tz) if enabled else None,
+    )
+    return types.SimpleNamespace(operating_hours=oh)
+
+
+def _fake_canary(paused=False):
+    return types.SimpleNamespace(is_paused=paused)
+
+
+@pytest.mark.parametrize(
+    "local_dt,expected",
+    [
+        # Monday 09:30 NY — exact open → active (None)
+        (_dt.datetime(2026, 4, 20, 9, 30), None),
+        # Monday 16:00 NY — exact close → inactive (>= stop)
+        (_dt.datetime(2026, 4, 20, 16, 0), "out_of_window_hours"),
+        # Monday 08:00 NY — before open
+        (_dt.datetime(2026, 4, 20, 8, 0), "out_of_window_hours"),
+        # Monday 12:00 NY — active
+        (_dt.datetime(2026, 4, 20, 12, 0), None),
+        # Saturday 12:00 NY — weekend
+        (_dt.datetime(2026, 4, 18, 12, 0), "out_of_window_weekend"),
+        # Sunday 23:00 NY — weekend
+        (_dt.datetime(2026, 4, 19, 23, 0), "out_of_window_weekend"),
+    ],
+)
+def test_operating_hours_skip_matrix(local_dt, expected):
+    """Timezone-aware start/stop + weekday checks."""
+    import atm.main as _main
+
+    cfg = _oh_cfg()
+    tz = cfg.operating_hours._tz_cache
+    now_ts = local_dt.replace(tzinfo=tz).timestamp()
+
+    lifecycle = _main.LifecycleState()
+    result = _main._should_skip(now_ts, lifecycle, cfg, _fake_canary())
+    assert result == expected
+
+
+def test_market_open_close_transitions_logged_once():
+    """Crossing a boundary emits exactly one market_open / market_closed event."""
+    import atm.main as _main
+
+    audit_events = []
+    alerts = []
+
+    class _A:
+        def log(self, e): audit_events.append(e)
+
+    class _N:
+        def send(self, a): alerts.append(a)
+
+    cfg = _oh_cfg()
+    tz = cfg.operating_hours._tz_cache
+    lifecycle = _main.LifecycleState()
+    canary = _fake_canary()
+
+    # Prime as closed (before open, Monday 08:00)
+    pre_open = _dt.datetime(2026, 4, 20, 8, 0, tzinfo=tz).timestamp()
+    skip_pre = _main._should_skip(pre_open, lifecycle, cfg, canary)
+    _main._maybe_log_transition(skip_pre, lifecycle, pre_open, _A(), _N())
+    # First evaluation seeds state, no alert yet.
+    assert lifecycle.last_window_state == "closed"
+    assert alerts == []
+    assert audit_events == []
+
+    # Transition to open
+    mid = _dt.datetime(2026, 4, 20, 12, 0, tzinfo=tz).timestamp()
+    skip_mid = _main._should_skip(mid, lifecycle, cfg, canary)
+    _main._maybe_log_transition(skip_mid, lifecycle, mid, _A(), _N())
+    assert lifecycle.last_window_state == "open"
+    assert len(alerts) == 1
+    assert any(e.get("event") == "market_open" for e in audit_events)
+
+    # Repeated open tick — no duplicate log
+    alerts.clear()
+    audit_events.clear()
+    skip_mid2 = _main._should_skip(mid + 60, lifecycle, cfg, canary)
+    _main._maybe_log_transition(skip_mid2, lifecycle, mid + 60, _A(), _N())
+    assert alerts == []
+    assert audit_events == []
+
+    # Transition to close
+    close = _dt.datetime(2026, 4, 20, 17, 0, tzinfo=tz).timestamp()
+    skip_close = _main._should_skip(close, lifecycle, cfg, canary)
+    _main._maybe_log_transition(skip_close, lifecycle, close, _A(), _N())
+    assert lifecycle.last_window_state == "closed"
+    assert any(e.get("event") == "market_closed" for e in audit_events)
+
+
+def test_market_transition_sends_notification():
+    """market_open / market_closed transitions produce kind=status alerts."""
+    import atm.main as _main
+
+    alerts = []
+
+    class _A:
+        def log(self, e): pass
+
+    class _N:
+        def send(self, a): alerts.append(a)
+
+    cfg = _oh_cfg()
+    tz = cfg.operating_hours._tz_cache
+    lifecycle = _main.LifecycleState(last_window_state="closed")
+
+    mid = _dt.datetime(2026, 4, 20, 12, 0, tzinfo=tz).timestamp()
+    _main._maybe_log_transition(None, lifecycle, mid, _A(), _N())
+    assert len(alerts) == 1
+    assert alerts[0].kind == "status"
+    assert "market" in alerts[0].title.lower() or "piața" in alerts[0].body.lower()
+
+
+def test_startup_in_window_suppresses_market_open():
+    """R2 #20: first evaluation in-window just seeds state; no alert fires."""
+    import atm.main as _main
+
+    alerts = []
+    events = []
+
+    class _A:
+        def log(self, e): events.append(e)
+
+    class _N:
+        def send(self, a): alerts.append(a)
+
+    cfg = _oh_cfg()
+    tz = cfg.operating_hours._tz_cache
+    lifecycle = _main.LifecycleState()  # last_window_state is None
+
+    in_window = _dt.datetime(2026, 4, 20, 12, 0, tzinfo=tz).timestamp()
+    skip = _main._should_skip(in_window, lifecycle, cfg, _fake_canary())
+    assert skip is None
+    _main._maybe_log_transition(skip, lifecycle, in_window, _A(), _N())
+
+    # Seeded silently
+    assert lifecycle.last_window_state == "open"
+    assert alerts == []
+    assert not any(e.get("event") == "market_open" for e in events)
+
+    # Two more ticks, still in-window → no spurious alert
+    for _ in range(2):
+        skip = _main._should_skip(in_window + 60, lifecycle, cfg, _fake_canary())
+        _main._maybe_log_transition(skip, lifecycle, in_window + 60, _A(), _N())
+    assert alerts == []
+
+
+def test_operating_hours_weekday_locale_independent():
+    """R2 #22: weekday check must not depend on process locale (strftime('%a'))."""
+    import locale as _locale
+    import atm.main as _main
+
+    cfg = _oh_cfg()
+    tz = cfg.operating_hours._tz_cache
+    # Saturday 12:00 NY
+    sat = _dt.datetime(2026, 4, 18, 12, 0, tzinfo=tz).timestamp()
+
+    original = _locale.setlocale(_locale.LC_TIME)
+    try:
+        for loc in ("C", "de_DE.UTF-8"):
+            try:
+                _locale.setlocale(_locale.LC_TIME, loc)
+            except _locale.Error:
+                continue  # locale not installed → skip gracefully
+            lifecycle = _main.LifecycleState()
+            result = _main._should_skip(sat, lifecycle, cfg, _fake_canary())
+            assert result == "out_of_window_weekend", (
+                f"locale={loc} returned {result!r}"
+            )
+    finally:
+        try:
+            _locale.setlocale(_locale.LC_TIME, original)
+        except _locale.Error:
+            _locale.setlocale(_locale.LC_TIME, "C")
+
+
+def test_should_skip_user_paused_wins():
+    import atm.main as _main
+    cfg = _oh_cfg()
+    lifecycle = _main.LifecycleState(user_paused=True)
+    # Mid-Monday (in-window) — should still skip because user_paused
+    tz = cfg.operating_hours._tz_cache
+    mid = _dt.datetime(2026, 4, 20, 12, 0, tzinfo=tz).timestamp()
+    assert _main._should_skip(mid, lifecycle, cfg, _fake_canary()) == "user_paused"
+
+
+def test_should_skip_canary_drift_wins_over_window():
+    import atm.main as _main
+    cfg = _oh_cfg()
+    lifecycle = _main.LifecycleState()
+    tz = cfg.operating_hours._tz_cache
+    mid = _dt.datetime(2026, 4, 20, 12, 0, tzinfo=tz).timestamp()
+    assert _main._should_skip(mid, lifecycle, cfg, _fake_canary(paused=True)) == "drift_paused"
+
+
+# ---------------------------------------------------------------------------
+# Commit 5: /pause /resume dispatch (plan tests #11-15, #16, R2 #21)
+# ---------------------------------------------------------------------------
+
+def _dispatch_ctx(canary=None, lifecycle=None, cfg=None):
+    """Minimal RunContext for _dispatch_command unit tests."""
+    import atm.main as _main
+
+    class _A:
+        def __init__(self): self.events = []
+        def log(self, e): self.events.append(e)
+
+    class _N:
+        def __init__(self): self.alerts = []
+        def send(self, a): self.alerts.append(a)
+
+    class _S:
+        is_running = False
+        interval_s = None
+        def start(self, s): self.is_running = True
+        def stop(self): self.is_running = False
+
+    if canary is None:
+        canary = types.SimpleNamespace(is_paused=False, resume=lambda: None)
+    if lifecycle is None:
+        lifecycle = _main.LifecycleState()
+    if cfg is None:
+        cfg = MagicMock()
+        cfg.telegram.auto_poll_interval_s = 180
+        cfg.operating_hours = types.SimpleNamespace(enabled=False, _tz_cache=None)
+
+    state = _main._LoopState(start=0.0)
+    ctx = _main.RunContext(
+        cfg=cfg, capture=lambda: None, canary=canary,
+        detector=MagicMock(), fsm=MagicMock(),
+        notifier=_N(), audit=_A(), detection_log=_A(),
+        scheduler=_S(), samples_dir=Path("."), fires_dir=Path("."),
+        cmd_queue=MagicMock(), state=state,
+        levels_extractor_factory=lambda *a, **kw: None,
+        lifecycle=lifecycle,
+    )
+    return ctx
+
+
+@pytest.mark.asyncio
+async def test_pause_command_sets_user_paused_and_skips_detection():
+    import atm.main as _main
+    from atm.commands import Command
+
+    ctx = _dispatch_ctx()
+    await _main._dispatch_command(ctx, Command(action="pause"))
+
+    assert ctx.lifecycle.user_paused is True
+    # When combined with _should_skip, we get user_paused
+    assert _main._should_skip(0.0, ctx.lifecycle, ctx.cfg, ctx.canary) == "user_paused"
+    # Audit + notif
+    assert any(e.get("event") == "user_paused" for e in ctx.audit.events)
+    assert any(a.kind == "status" and "oprit" in a.title.lower() for a in ctx.notifier.alerts)
+
+
+@pytest.mark.asyncio
+async def test_resume_clears_user_paused_and_canary_when_forced():
+    import atm.main as _main
+    from atm.commands import Command
+
+    canary_state = {"paused": True}
+    canary = types.SimpleNamespace(
+        is_paused=True,
+        resume=lambda: canary_state.__setitem__("paused", False),
+    )
+    # Re-bind is_paused via property so resume() effect is visible
+    class _Canary:
+        def __init__(self): self._p = True
+        @property
+        def is_paused(self): return self._p
+        def resume(self): self._p = False
+    canary = _Canary()
+
+    ctx = _dispatch_ctx(canary=canary)
+    ctx.lifecycle.user_paused = True
+
+    await _main._dispatch_command(ctx, Command(action="resume", value=1))
+
+    assert ctx.lifecycle.user_paused is False
+    assert canary.is_paused is False
+    force_events = [e for e in ctx.audit.events if e.get("event") == "user_resumed"]
+    assert force_events and force_events[0]["force"] is True
+
+
+@pytest.mark.asyncio
+async def test_resume_during_drift_keeps_canary_paused_without_force():
+    """R2 #21: plain /resume during drift clears user_paused but NOT canary."""
+    import atm.main as _main
+    from atm.commands import Command
+
+    class _Canary:
+        def __init__(self): self._p = True
+        @property
+        def is_paused(self): return self._p
+        def resume(self): self._p = False
+    canary = _Canary()
+
+    ctx = _dispatch_ctx(canary=canary)
+    ctx.lifecycle.user_paused = True
+
+    await _main._dispatch_command(ctx, Command(action="resume"))  # no force
+
+    assert ctx.lifecycle.user_paused is False
+    assert canary.is_paused is True  # still drift-paused
+    # Message must mention drift
+    status = [a for a in ctx.notifier.alerts if a.kind == "status"]
+    assert status and ("drift" in (status[0].title + status[0].body).lower())
+
+    # Now force
+    ctx.notifier.alerts.clear()
+    await _main._dispatch_command(ctx, Command(action="resume", value=1))
+    assert canary.is_paused is False
+
+
+@pytest.mark.asyncio
+async def test_resume_out_of_window_responds_with_pending_message():
+    """/resume while operating-hours window is closed → special body."""
+    import atm.main as _main
+    from atm.commands import Command
+
+    cfg = _oh_cfg()
+    tz = cfg.operating_hours._tz_cache
+    lifecycle = _main.LifecycleState(user_paused=True, last_window_state="closed")
+    canary = types.SimpleNamespace(is_paused=False, resume=lambda: None)
+
+    ctx = _dispatch_ctx(canary=canary, lifecycle=lifecycle, cfg=cfg)
+
+    # Pin time to Saturday
+    import atm.main as _mm
+    real_time = _mm.time
+    fake_ts = _dt.datetime(2026, 4, 18, 12, 0, tzinfo=tz).timestamp()
+    class _FakeTime:
+        def time(self): return fake_ts
+        def monotonic(self): return 0.0
+    _mm.time = _FakeTime()
+    try:
+        await _main._dispatch_command(ctx, Command(action="resume"))
+    finally:
+        _mm.time = real_time
+
+    assert ctx.lifecycle.user_paused is False
+    status = [a for a in ctx.notifier.alerts if a.kind == "status"]
+    assert status
+    combined = (status[0].title + status[0].body).lower()
+    assert "închis" in combined or "piața" in combined or "ferestr" in combined
+
+
+@pytest.mark.asyncio
+async def test_status_command_reports_pause_reason():
+    """/status body must mention pause reason + window state."""
+    import atm.main as _main
+    from atm.commands import Command
+
+    ctx = _dispatch_ctx()
+    ctx.lifecycle.user_paused = True
+    # Stub detector.rolling for status
+    ctx.detector.rolling = []
+    ctx.fsm.state = types.SimpleNamespace(value="IDLE")
+
+    await _main._dispatch_command(ctx, Command(action="status"))
+
+    status = [a for a in ctx.notifier.alerts if a.kind == "status"]
+    assert status
+    body = status[0].body
+    assert "user_paused" in body or "pauzat:user_paused" in body
+
+
+@pytest.mark.asyncio
+async def test_lifecycle_with_drift_then_resume_then_fire(monkeypatch, tmp_path):
+    """E2E #16: drift paused → /resume force → dark_red/light_red produce FIRE alert.
+
+    This test verifies the full command-driven lifecycle in isolation:
+    - canary starts drift-paused, _should_skip returns drift_paused
+    - /resume force clears canary + user_paused
+    - subsequent detection produces SELL fire through normal FSM path
+    """
+    import atm.main as _main
+    from atm.commands import Command
+
+    # Canary with mutable pause state
+    class _Canary:
+        def __init__(self): self._p = True
+        @property
+        def is_paused(self): return self._p
+        def resume(self): self._p = False
+
+    canary = _Canary()
+    cfg = MagicMock()
+    cfg.telegram.auto_poll_interval_s = 180
+    cfg.operating_hours = types.SimpleNamespace(enabled=False, _tz_cache=None)
+
+    ctx = _dispatch_ctx(canary=canary, cfg=cfg)
+
+    # 1. While drift-paused, _should_skip returns drift_paused
+    assert _main._should_skip(0.0, ctx.lifecycle, cfg, canary) == "drift_paused"
+
+    # 2. User issues /resume force
+    await _main._dispatch_command(ctx, Command(action="resume", value=1))
+    assert canary.is_paused is False
+    assert _main._should_skip(0.0, ctx.lifecycle, cfg, canary) is None
+
+    # 3. Feed a yellow→light_red sequence through _handle_tick (FSM path)
+    from atm.state_machine import StateMachine, State
+    fsm = StateMachine(lockout_s=60)
+
+    class _N:
+        def __init__(self): self.alerts = []
+        def send(self, a): self.alerts.append(a)
+
+    class _A:
+        def log(self, _e): pass
+
+    notif = _N()
+    audit = _A()
+    cfg_mock = types.SimpleNamespace(alerts=types.SimpleNamespace(fire_on_phase_skip=True))
+
+    _main._handle_tick(fsm, "yellow", 1.0, notif, audit, first_accepted=False, cfg=cfg_mock)
+    _main._handle_tick(fsm, "dark_red", 2.0, notif, audit, first_accepted=False, cfg=cfg_mock)
+    tr = _main._handle_tick(fsm, "light_red", 3.0, notif, audit, first_accepted=False, cfg=cfg_mock)
+
+    # FSM reached fire via normal path
+    assert tr is not None and tr.trigger == "SELL"
+    assert fsm.state == State.IDLE
diff --git a/tests/test_validate.py b/tests/test_validate.py
new file mode 100644
index 0000000..c8e8dfc
--- /dev/null
+++ b/tests/test_validate.py
@@ -0,0 +1,214 @@
+"""Tests for atm.validate — offline calibration validation.
+
+Covers the 3 tests from plan section D':
+  17. test_validate_calibration_pass
+  18. test_validate_calibration_fail_reports_top_candidates
+  19. test_validate_calibration_file_not_found
+"""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+from atm.config import (
+    CanaryRegion,
+    ColorSpec,
+    Config,
+    DiscordCfg,
+    ROI,
+    TelegramCfg,
+    YAxisCalib,
+)
+from atm.detector import DetectionResult
+from atm.vision import ColorMatch
+
+
+def _make_config() -> Config:
+    """Minimal Config with a palette large enough to support top-3 candidates."""
+    colors = {
+        "turquoise":   ColorSpec(rgb=(0, 200, 200),    tolerance=30),
+        "yellow":      ColorSpec(rgb=(255, 255, 0),    tolerance=30),
+        "dark_green":  ColorSpec(rgb=(0, 100, 0),      tolerance=30),
+        "dark_red":    ColorSpec(rgb=(165, 42, 42),    tolerance=30),
+        "light_green": ColorSpec(rgb=(144, 238, 144),  tolerance=30),
+        "light_red":   ColorSpec(rgb=(255, 182, 193),  tolerance=30),
+        "gray":        ColorSpec(rgb=(128, 128, 128),  tolerance=30),
+        "background":  ColorSpec(rgb=(18, 18, 18),     tolerance=15),
+    }
+    return Config(
+        window_title="test",
+        dot_roi=ROI(x=0, y=0, w=100, h=100),
+        chart_roi=ROI(x=0, y=0, w=100, h=100),
+        colors=colors,
+        y_axis=YAxisCalib(p1_y=0, p1_price=100.0, p2_y=100, p2_price=0.0),
+        canary=CanaryRegion(
+            roi=ROI(x=0, y=0, w=10, h=10),
+            baseline_phash="0" * 64,
+        ),
+        discord=DiscordCfg(webhook_url="http://localhost/fake"),
+        telegram=TelegramCfg(bot_token="fake_token", chat_id="123"),
+        debounce_depth=1,
+    )
+
+
+def _write_labels(tmp_path: Path, entries: list[dict]) -> Path:
+    f = tmp_path / "labels.json"
+    f.write_text(json.dumps(entries), encoding="utf-8")
+    return f
+
+
+def _write_blank_png(tmp_path: Path, name: str) -> Path:
+    """Write a trivially-valid 10x10 BGR image so cv2.imread returns non-None."""
+    import cv2
+    p = tmp_path / name
+    arr = np.zeros((10, 10, 3), dtype=np.uint8)
+    cv2.imwrite(str(p), arr)
+    return p
+
+
+# ---------------------------------------------------------------------------
+# Test 17: PASS path — mocked Detector.step returns expected color
+# ---------------------------------------------------------------------------
+
+def test_validate_calibration_pass(monkeypatch, tmp_path):
+    from atm import validate as validate_mod
+
+    img_path = _write_blank_png(tmp_path, "yellow_sample.png")
+    labels = _write_labels(
+        tmp_path,
+        [{"path": str(img_path), "expected": "yellow", "note": "test"}],
+    )
+
+    def fake_step(self, ts, frame=None):
+        return DetectionResult(
+            ts=ts,
+            window_found=True,
+            dot_found=True,
+            rgb=(250, 250, 5),
+            match=ColorMatch(name="yellow", distance=6.0, confidence=0.94),
+            accepted=True,
+            color="yellow",
+        )
+
+    monkeypatch.setattr("atm.detector.Detector.step", fake_step)
+
+    report = validate_mod.validate_calibration(labels, _make_config())
+
+    assert report.total == 1
+    assert report.passed == 1
+    assert report.failed == 0
+    assert report.all_pass is True
+    rec = report.records[0]
+    assert rec.passed is True
+    assert rec.detected == "yellow"
+    assert rec.expected == "yellow"
+    assert "[PASS]" in report.render()
+
+    # CLI wiring: exit 0
+    import atm.main as _main
+
+    class _Args:
+        label_file = labels
+
+    monkeypatch.setattr("atm.config.Config.load_current", classmethod(lambda cls, d: _make_config()))
+    with pytest.raises(SystemExit) as exc_info:
+        _main._cmd_validate_calibration(_Args())
+    assert exc_info.value.code == 0
+
+
+# ---------------------------------------------------------------------------
+# Test 18: FAIL path — Detector returns wrong color; report lists top 3
+#                      candidates and a SUGGESTIONS line with RGB distance.
+# ---------------------------------------------------------------------------
+
+def test_validate_calibration_fail_reports_top_candidates(monkeypatch, tmp_path):
+    from atm import validate as validate_mod
+
+    img_path = _write_blank_png(tmp_path, "dark_red_sample.png")
+    labels = _write_labels(
+        tmp_path,
+        [{"path": str(img_path), "expected": "dark_red", "note": "missed dark_red"}],
+    )
+
+    # Observed RGB closer to gray than dark_red (like the real 2026-04-17 miss).
+    def fake_step(self, ts, frame=None):
+        return DetectionResult(
+            ts=ts,
+            window_found=True,
+            dot_found=True,
+            rgb=(135, 62, 67),
+            match=ColorMatch(name="gray", distance=45.0, confidence=0.12),
+            accepted=True,
+            color="gray",
+        )
+
+    monkeypatch.setattr("atm.detector.Detector.step", fake_step)
+
+    report = validate_mod.validate_calibration(labels, _make_config())
+
+    assert report.total == 1
+    assert report.failed == 1
+    assert report.all_pass is False
+
+    rec = report.records[0]
+    assert rec.passed is False
+    assert rec.detected == "gray"
+    assert rec.expected == "dark_red"
+    # Top 3 candidates populated (name, score) sorted by RGB distance.
+    assert len(rec.top3) == 3
+    names = [n for n, _ in rec.top3]
+    # dark_red should appear in top candidates since observed RGB(135,62,67)
+    # is reasonably close to dark_red(165,42,42).
+    assert "dark_red" in names
+
+    rendered = report.render()
+    assert "[FAIL]" in rendered
+    assert "Top 3 candidates:" in rendered
+    assert "SUGGESTIONS:" in rendered
+    # The suggestion must mention the expected color's RGB and the measured distance.
+    assert "dark_red" in rendered
+    assert "(165, 42, 42)" in rendered
+
+    # CLI wiring: exit 1
+    import atm.main as _main
+
+    class _Args:
+        label_file = labels
+
+    monkeypatch.setattr("atm.config.Config.load_current", classmethod(lambda cls, d: _make_config()))
+    with pytest.raises(SystemExit) as exc_info:
+        _main._cmd_validate_calibration(_Args())
+    assert exc_info.value.code == 1
+
+
+# ---------------------------------------------------------------------------
+# Test 19: missing label file — clean error, non-zero exit, no stack trace
+# ---------------------------------------------------------------------------
+
+def test_validate_calibration_file_not_found(monkeypatch, tmp_path, capsys):
+    from atm import validate as validate_mod
+
+    missing = tmp_path / "nope.json"
+
+    # Library-level: raises ValidationError (not bare FileNotFoundError).
+    with pytest.raises(validate_mod.ValidationError) as exc_info:
+        validate_mod.validate_calibration(missing, _make_config())
+    assert "not found" in str(exc_info.value).lower()
+
+    # CLI-level: graceful sys.exit with non-zero code, message on stderr.
+    import atm.main as _main
+
+    class _Args:
+        label_file = missing
+
+    monkeypatch.setattr("atm.config.Config.load_current", classmethod(lambda cls, d: _make_config()))
+    with pytest.raises(SystemExit) as exc_info:
+        _main._cmd_validate_calibration(_Args())
+    assert exc_info.value.code != 0
+    err = capsys.readouterr().err
+    assert "not found" in err.lower()
+    # Ensure no python traceback leaked through.
+    assert "Traceback" not in err