feat(vision): state the box extent in the grounding prompts (gate item 3)

[AlexSKuznetsov]

What

Gate item 3 from docs/19-Post-Spike-Build-Plan.md. Every prompt that emits a box which then crops an object for embedding now carries one extent rule naming what to include — the whole object a person would point to (a keyring's ring/fob/lanyard, a mug's handle, glasses' arms), not its single most recognizable part.

Applied in reason/cosmos.py to:

• _PROMPT — the event grounder (live query path, pipeline.py:290)
• _LOCALIZE_PROMPT — the semantic_box re-grounding (also live, pipeline.py:387)
• _TEMPORAL_LOCALIZE_PROMPT — enrollment

_REFERENCE_PROMPT emits no box (VALID/REJECT) and is untouched.

Why

• Spike 12c: naming the extent takes the worst noun from grounding IoU 0.12 → 0.92 and kills a temperature-0 bimodal swing.
• Spike 3c: the shipped wording conflated validity with extent and pinned the keys crop to the near-generic metal blade, costing identity F1 0.776 → 0.939 on the same images/backbone.

So the keys requirement now separates the two jobs: a metal key blade must be present for validity (keyboard still rejects to NO_OBJECT), but the box covers the whole keyring. The localizers' exclusion lists stop excluding the attached cord/lanyard/strap they should now include, keeping only hand/screen/background. The extent rule's closing clause ("never … the whole frame") guards the failure that held an earlier attempt, where "box the whole object" was read as "box everything" and re-admitted the keyboard.

Prompt versions bumped (reason-v2, localize-v4, temporal-v2) so recorded runs carry the wording.

Tests

• Golden-prompt test asserts the extent rule is present on all three box-emitting prompts.
• A second asserts the keys box is the whole keyring, not the blade alone.
• 174 passed, 2 deselected; ruff + pyright clean.

Local validation (mechanism), and its limits

Ran a real LFM2.5-VL locally (native grounding prompt ± a short extent clause) over the 36 keys probe stills, measuring box-area fraction of frame (EXTENT.md's metric; no ground-truth boxes locally):

	1.6B	3B (spike's graded model)
median box-area, native → +extent	0.072 → 0.101 (+40%)	0.099 → 0.113 (+14%)
widened / shrank / same (of 36)	29 / 5 / 2	19 / 15 / 2

In aggregate the extent sentence enlarges the box on both models — the item-3 mechanism. On the 3B it is bimodal per-image (19 up / 15 down), which independently reproduces the temperature-0 swing spike 12c flagged and said the validated wording is what kills.

Caveats (deliberate): the local run used a paraphrase clause on LFM's native prompt, not the committed production _EXTENT_RULE on the Cosmos-format prompt; and box-area is not IoU. The graded 0.12 → 0.92 IoU was measured on hosted Gemini at temp 0 and needs the validated prompt + ground truth on the H200 target to reproduce — not this laptop. The keyboard-rejection guard is likewise unverifiable locally (LFM never emits NO_OBJECT, per spike 3c).

🤖 Generated with Claude Code