fix(inference): stop polluting schema-mode crop-describe responses

[Liuhaai]

Summary

/api/inference/crop-describe was silently mutating the wire response when callers passed a response_format=json_schema. Two bugs combined to make cortex's vehicle make field empty even when the VLM clearly recognized the make in its prose summary.

The bugs

1. _normalize_entity_item injected schema-forbidden fields

For every vehicle entity the normalizer ran:

label = norm.get("brand") or norm.get("make") or norm.get("description", "")
norm.setdefault("brand", label)
norm.setdefault("make", label)

When the model emitted {"id":"nv0","type":"suv","make":""}:

• norm.get("make") returns "" (falsy) → falls through to description → ""
• setdefault("brand", "") adds a brand key the schema never declared

Observed in prod (cam_1dd58540, frame obs_74a6bc82052b): the VLM summary said "A blue Audi SUV is parked near a payment kiosk…" yet the structured vehicles[] arrived at cortex as {"id":"nv0","type":"suv","make":"","brand":""}. The schema's additionalProperties: False contract was bypassed post-VLM.

2. Duplicate YOLO-context auto-prepend

Cortex already enumerates per-id detections in its scene_prompt (DETECTIONS:\n id=nv0 vehicle bbox=...). trio-core matched on the literal substring "YOLO detections" (which cortex's prompt does not contain) and prepended a second YOLO bbox table ahead of the caller's prompt.

Empirical A/B on qwen3-vl-flash with the Audi frame:

• Cortex's prompt alone → make: "Audi" in 3/3 trials
• Same prompt + trio-core's duplicate YOLO prefix → make: "Audi" in only 1/3 trials; the other 2/3 collapsed to make: "unknown"

Fix

Gate both behaviors on req.response_format is None:

• _normalize_entities(entities, schema_mode=True) returns the parsed dict unchanged — no field invention, no iteration over animals/keys the caller's schema didn't declare.
• The YOLO-context auto-prepend is skipped when a schema is in effect.

The legacy free-text default scene_prompt path keeps the string-to-dict fallback and the YOLO hint prepend; existing tests cover that and a new test pins it down.

Test plan

• pytest tests/test_inference_router.py — 15 passed
• Full suite: 419 passed, 7 skipped (no regressions)
• New regression tests:
  • test_normalize_entities_schema_mode_preserves_empty_make
  • test_normalize_entities_schema_mode_drops_extra_keys_in_passthrough
  • test_crop_describe_schema_mode_skips_yolo_context_prepend
  • test_crop_describe_default_prompt_still_gets_yolo_context (legacy path preserved)
• Manual A/B against the prod Audi frame via DashScope confirms duplicate-context fix restores attribute extraction.

🤖 Generated with Claude Code