fix(memory): retry cu split la eroare de context Ollama in indexare KB

Chunk-urile mari (tabele dense cu IP-uri/cod) treceau de split-ul pe
paragrafe dar depaseau limita reala de tokeni a all-minilm, nu doar
un prag fix de caractere — 18/20 chunk-uri din infrastructure.md
esuau silentios la fiecare heartbeat. embed_chunk() acum injumatateste
textul recursiv la eroare 500 si retrimite pana reuseste.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

src/memory_search.py

@@ -113,6 +113,28 @@ def get_embedding(text: str) -> list[float]:
         raise ConnectionError(f"Ollama API error: {e.response.status_code}")
 
 
+def embed_chunk(text: str, max_depth: int = 6) -> list[tuple[str, list[float]]]:
+    """Embed text, halving on Ollama context-length errors until it fits.
+
+    The model's context limit is token-based, not character-based, so dense
+    content (tables, IPs, code) can overflow well under _CHUNK_MAX chars. This
+    self-corrects instead of guessing a safe static threshold.
+    """
+    try:
+        return [(text, get_embedding(text))]
+    except ConnectionError:
+        if max_depth <= 0 or len(text) < _CHUNK_MIN:
+            raise
+        mid = len(text) // 2
+        split_at = text.rfind("\n", 0, mid)
+        if split_at <= 0:
+            split_at = mid
+        left, right = text[:split_at].strip(), text[split_at:].strip()
+        if not left or not right:
+            raise
+        return embed_chunk(left, max_depth - 1) + embed_chunk(right, max_depth - 1)
+
+
 def serialize_embedding(embedding: list[float]) -> bytes:
     """Pack floats to bytes for SQLite storage."""
     return struct.pack(f"{len(embedding)}f", *embedding)
@@ -194,15 +216,17 @@ def index_file(file_path: Path) -> int:
     conn = get_db()
     try:
         conn.execute("DELETE FROM chunks WHERE file_path = ?", (rel_path,))
-        for i, chunk_text in enumerate(chunks):
-            embedding = get_embedding(chunk_text)
+        i = 0
+        for chunk_text in chunks:
+            for sub_text, embedding in embed_chunk(chunk_text):
                 conn.execute(
                     """INSERT INTO chunks (file_path, chunk_index, chunk_text, embedding, updated_at)
                        VALUES (?, ?, ?, ?, ?)""",
-                (rel_path, i, chunk_text, serialize_embedding(embedding), now),
+                    (rel_path, i, sub_text, serialize_embedding(embedding), now),
                 )
+                i += 1
         conn.commit()
-        return len(chunks)
+        return i
     finally:
         conn.close()