fix(inference): stop dumping raw JSON into description on parse failure

[Liuhaai]

Summary

When the VLM emits a SCENE_SCHEMA-shaped JSON but json.loads fails (truncation, no closing brace, or malformed), _crop_describe_inner was leaking a 300-char raw JSON slice into description — observable as DB rows like:

{
  "activity_level": "moderate",
  "people_count": 0,
  ...
  "is_know

(ending mid-token at exactly 300 chars). The downstream substring-match fallback then guessed activity_level from random words inside the JSON dump, producing rows where the structured columns were inferred from JSON literals rather than the model's actual output.

Changes

• Recognize lowercase summary (the SCENE_SCHEMA contract) in the entities branch so schema-mode outputs produce a clean description instead of the "0 people, N vehicles: parked, …" parts template. summary is read without pop so callers re-parsing entities still see it.
• Stop dumping raw JSON in the final fallback. When json.loads fails, or when text starts with { but has no closing } (truncation), return desc = "" instead of clean[:300].
• Add diagnostic warnings with text length + head/tail bytes so the truncation-vs-malformed root cause is observable from logs once deployed.
• Plain prose (no leading {) keeps the existing 300-char slice behavior unchanged.

Test plan

• All 6 existing tests in test_inference_router.py pass unchanged.
• New: SCENE_SCHEMA summary populates description, summary preserved in entities.
• New: truncated JSON (no closing brace) → desc="", warning logged with truncation suspected.
• New: malformed JSON (trailing comma) → desc="", warning logged with JSON parse failed.
• New: plain prose still uses the 300-char slice.
• After deploy: confirm crop_describe: JSON parse failed / truncation suspected warnings appear in /tmp/vlm.log and pin the root cause (likely max_tokens=512 default for SCENE_SCHEMA payloads with multi-vehicle arrays).

Notes

This addresses the symptom only. The next step is using the new diagnostic logs to determine the failure mode and either bump max_tokens (currently engine.config.max_tokens=512 and not overridden in _crop_describe_inner) or fix the specific malformation. ~500 existing JSON-string description rows in prod are not backfilled by this change.

🤖 Generated with Claude Code