feat(diarization-ui): dynamic num_ctx for Ollama based on prompt size

Estimates required context window from prompt length (chars/3 + 2048
response buffer) and rounds up to the nearest fixed tier (4096, 8192,
16384, 32768, 65536). Fixed tiers prevent Ollama from reloading the
model on every call. The chosen num_ctx is prepended to the stored
llm_prompt so it's visible in the debug window.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -36,6 +36,14 @@ def now_iso() -> str:
     return datetime.utcnow().isoformat()
 
 
+def _estimate_num_ctx(prompt: str) -> int:
+    needed = len(prompt) // 3 + 2048  # rough token estimate + response buffer
+    for ctx in (4096, 8192, 16384, 32768, 65536):
+        if needed <= ctx:
+            return ctx
+    return 65536
+
+
 def init_db():
     os.makedirs(os.path.dirname(DB_PATH), exist_ok=True)
     with db() as c:
@@ -371,13 +379,14 @@ def _process_analysis_job(job_id: int):
             + f"\\nTEXT:\\n{doc['content_md']}\\n"
         )
 
-        _job_set(job_id, llm_prompt=llm_prompt)
+        num_ctx = _estimate_num_ctx(llm_prompt)
+        _job_set(job_id, llm_prompt=f"[num_ctx={num_ctx}]\n\n{llm_prompt}")
         with _JOB_STREAM_LOCK:
             _JOB_STREAMS[job_id] = {"thinking": "", "response": ""}
 
         r = requests.post(
             f"{OLLAMA_BASE_URL}/api/generate",
-            json={"model": OLLAMA_MODEL, "prompt": llm_prompt, "stream": True},
+            json={"model": OLLAMA_MODEL, "prompt": llm_prompt, "stream": True, "options": {"num_ctx": num_ctx}},
             stream=True,
             timeout=1200,
         )