fix: parse_turn falls back gracefully + classify logs flapping classifiers

The turn endpoint was 500ing in multi-bot scenes whenever the
classifier provider hiccuped on parse_turn — particularly visible
after a guest was added and bots started exchanging turns. The
traceback was 'classify failed for schema ParsedTurn with no default'
because parse_turn was the only classify caller without a default.

Two changes:

- chat/services/turn_parse.py: parse_turn now passes a default that
  wraps the whole prose as one 'dialogue' segment. The narrative
  still fires on the prose; we lose finer-grained segment kinds
  (action vs dialogue vs ooc) on this turn, but the request returns
  cleanly. Updated the existing test that pinned the old
  RuntimeError contract.

- chat/llm/classify.py: when retries are exhausted, log a WARNING
  with the schema name, last error type, and a snippet of the last
  raw text the model returned. Surfaces flapping classifiers in the
  uvicorn log for diagnosis without taking down the request.

Suite: 471 passed in 11.7s.

chat/llm/classify.py

@@ -1,11 +1,13 @@
 from __future__ import annotations
 import json
 import asyncio
+import logging
 from typing import TypeVar
 from pydantic import BaseModel, ValidationError
 from .client import LLMClient, Message
 
 T = TypeVar("T", bound=BaseModel)
+_log = logging.getLogger(__name__)
 
 REFUSAL_PATTERNS = ("i can't", "i cannot", "i'm sorry, but", "as an ai")
 
@@ -47,6 +49,8 @@ async def classify(
     # several seconds on tokens we'll never use. Classifier responses
     # are small JSON objects — 512 tokens is generous; usual completions
     # are 50-150.
+    last_text = None
+    last_error: BaseException | None = None
     for attempt in range(3):
         try:
             text = await asyncio.wait_for(
@@ -58,16 +62,29 @@ async def classify(
                 ),
                 timeout=timeout_s,
             )
+            last_text = text
             cleaned = _strip_json_fences(text)
             if any(p in cleaned.lower()[:80] for p in REFUSAL_PATTERNS) and not cleaned.lstrip().startswith("{"):
                 raise ValueError("refusal-shaped response")
             return schema.model_validate_json(cleaned)
-        except (ValidationError, ValueError, json.JSONDecodeError, asyncio.TimeoutError):
+        except (ValidationError, ValueError, json.JSONDecodeError, asyncio.TimeoutError) as exc:
+            last_error = exc
             msgs[0] = Message(
                 role="system",
                 content=system + schema_block + "\n\nRespond with valid JSON ONLY. No prose, no markdown fences.",
             )
             continue
+    # Log when we're falling back so flapping classifiers are
+    # diagnosable without taking down the request.
+    snippet = (last_text or "")[:200].replace("\n", " ")
+    _log.warning(
+        "classify(%s) exhausted 3 attempts; last_error=%s last_text=%r; "
+        "falling back to %s",
+        schema.__name__,
+        type(last_error).__name__ if last_error else "?",
+        snippet,
+        "default" if default is not None else "RuntimeError (no default)",
+    )
     if default is None:
         raise RuntimeError(f"classify failed for schema {schema.__name__} with no default")
     return default

chat/services/turn_parse.py

@@ -107,13 +107,23 @@ async def parse_turn(
     without an LLM call (the classifier would error on empty input
     anyway, and the result is unambiguous).
 
-    Raises ``RuntimeError`` if the classifier fails twice — no default
-    is supplied, since the caller (T19's turn flow) is responsible for
-    surfacing the error to the user.
+    Falls back to a single dialogue-shaped segment containing the
+    whole prose if the classifier flaps after retries — the turn flow
+    can keep moving (the narrative will still fire on the prose) at
+    the cost of finer-grained segment classification. The original
+    code raised ``RuntimeError`` here, which 500'd the whole request
+    and was particularly painful in multi-bot scenes where every
+    user turn paid the classifier round-trip.
     """
     if not prose.strip():
         return ParsedTurn(segments=[])
 
+    fallback = ParsedTurn(
+        segments=[TurnSegment(kind="dialogue", text=prose)],
+        intent="narrative",
+        landing_state_hint="",
+    )
+
     user_prompt = f"INPUT:\n{prose}"
     return await classify(
         client,
@@ -121,5 +131,6 @@ async def parse_turn(
         system=_SYSTEM_PROMPT,
         user=user_prompt,
         schema=ParsedTurn,
+        default=fallback,
         timeout_s=timeout_s,
     )

tests/test_turn_parse.py

@@ -73,11 +73,25 @@ async def test_parse_turn_empty_prose_short_circuits_without_classifier_call():
 
 
 @pytest.mark.asyncio
-async def test_parse_turn_raises_when_classifier_fails_twice():
+async def test_parse_turn_falls_back_to_whole_prose_when_classifier_fails():
+    """A flapping classifier (3 invalid responses) no longer 500s the
+    request. ``parse_turn`` returns the original prose as a single
+    ``dialogue`` segment so the turn flow can keep moving — the
+    narrative will still fire on the prose, just without finer-grained
+    segment classification.
+
+    The old contract was ``RuntimeError`` (no default), but in
+    production that took down the whole turn endpoint with a 500 the
+    moment any classifier provider hiccuped — particularly painful in
+    multi-bot scenes where every user turn pays the parse_turn cost.
+    """
     mock = MockLLMClient(canned=["nope", "still nope", "nope3"])
-    with pytest.raises(RuntimeError):
-        await parse_turn(
-            mock,
-            model="m",
-            prose='*shrugs* "whatever"',
-        )
+    result = await parse_turn(
+        mock,
+        model="m",
+        prose='*shrugs* "whatever"',
+    )
+    assert len(result.segments) == 1
+    assert result.segments[0].kind == "dialogue"
+    assert result.segments[0].text == '*shrugs* "whatever"'
+    assert result.intent == "narrative"