fix: classifier robustness — schema in prompt, retries, kickoff fallback

The kickoff parse-and-confirm route was 500-ing intermittently because
Hermes-3 + Featherless's response_format={"type":"json_object"} only
guarantees JSON output, NOT a particular schema. The model was inventing
its own field names (sceneTime, entities, settingDetails) instead of
the KickoffParse fields, causing Pydantic validation to fail on both
classify() retries.

Three changes:

1. Include the Pydantic JSON schema in the system prompt so the model
   knows exactly which keys to produce. Affects every classify() call
   (kickoff parse, turn parse, scene-close detect, significance,
   state-update, scene summarize). Strip ```json fences if the model
   wraps its output. Bump retries 2 → 3 (model is stochastic; one extra
   attempt closes most of the remaining gap).

2. parse_kickoff() now passes a default empty KickoffParse so the
   route degrades to a fillable form instead of 500 when the classifier
   ultimately fails. The confirm form is the human-in-the-loop; an
   empty form is strictly better UX than a stack trace.

3. Tests updated: bumped canned-failure arrays from 2 → 3 entries to
   match the new attempt count; renamed kickoff test from
   "raises_when_classifier_fails_twice" to
   "falls_back_to_empty_when_classifier_fails" reflecting the new
   degraded-but-usable behavior.

Verified live with all 3 sample bots (maya/eli/sam) — kickoff route
returns 200 across multiple attempts. Full suite: 168 passed.

chat/services/kickoff.py

@@ -85,6 +85,36 @@ def _build_user_prompt(
     )
 
 
+def _empty_activity() -> ActivityShape:
+    return ActivityShape(
+        posture="",
+        action_verb="",
+        action_interruptible=True,
+        action_required_attention="low",
+        action_expected_duration="brief",
+    )
+
+
+def _empty_kickoff_parse() -> KickoffParse:
+    """Default returned when the classifier can't produce a valid parse.
+
+    The user gets a mostly-empty confirm form they can fill in by hand
+    instead of a 500. ``initial_time_iso`` is left as the current UTC.
+    """
+    from datetime import datetime, timezone
+
+    return KickoffParse(
+        container_name="",
+        container_type="",
+        container_properties={},
+        you_activity=_empty_activity(),
+        bot_activity=_empty_activity(),
+        initial_time_iso=datetime.now(timezone.utc).isoformat(timespec="seconds"),
+        edge_seed_summary="",
+        edge_seed_knowledge_facts=[],
+    )
+
+
 async def parse_kickoff(
     client: LLMClient,
     *,
@@ -98,11 +128,9 @@ async def parse_kickoff(
 ) -> KickoffParse:
     """Parse authored kickoff prose into a structured ``KickoffParse``.
 
-    Internally calls :func:`chat.llm.classify.classify` with a labeled
-    user prompt. Raises ``RuntimeError`` if the classifier fails twice in
-    a row — no default is supplied at this layer, since the caller (T13's
-    confirm form) is responsible for showing an error and letting the
-    user edit.
+    Falls back to a mostly-empty default if the classifier fails — the
+    confirm-and-edit form is the human-in-the-loop, so a degraded form
+    that the user can fill in is preferable to a 500.
     """
     user_prompt = _build_user_prompt(
         bot_name=bot_name,
@@ -117,5 +145,6 @@ async def parse_kickoff(
         system=_SYSTEM_PROMPT,
         user=user_prompt,
         schema=KickoffParse,
+        default=_empty_kickoff_parse(),
         timeout_s=timeout_s,
     )