feat(identity): condition box padding on the box's source (gate item 4)

[AlexSKuznetsov]

What

Splits the single global identity_box_padding (0.12) into two source-conditioned knobs, applied at the call site that knows where the box came from:

Box source	Knob	Default	Call sites
Grounder (VLM analyze/localize)	identity_grounder_box_padding	0.75	query path (pipeline.py, ×3) + grounded enroll (enroll.py, ×2)
Tracker (SAM2, register button)	identity_tracker_box_padding	0.0	enroll_center_anchor (enroll.py, ×1)

enroll_manual is unaffected — it masks with a full-ones mask, so padding never applies there.

Why

Gate item 4 (docs/19-Post-Spike-Build-Plan.md), spikes 3e/2e. One padding value can't serve both sources:

• A grounder box can be cropped tight to the object's most recognizable part; widening it repairs the scoping (F1 0.776 → 0.909, spike 3e).
• A tracker box has no such defect, so the same padding only pulls in desk and hurts separation (+0.340 → +0.255, spike 2e).

Local validation (padding sweep on the identity probe)

Ran a real sweep to pick the grounder default rather than inherit 0.75 blind: LFM-3B grounds every probe image (local proxy for Cosmos), then the production crop path (box_to_mask → prepare_masked_crop) at each padding → C-RADIOv4-H → the probe's balanced accept/reject scoring. native = bare LFM box; extent = LFM + item 3's extent clause (post-item-3 proxy).

boxset    pad    sep     F1    AUC
native   0.12  0.018  0.432  0.546   ← old shipped default (too low)
native   0.75  0.115  0.824  0.875
extent   0.00  0.112  0.846  0.864
extent   0.50  0.163  0.917  0.942   ← best separation/margin
extent   0.75  0.153  0.936  0.949   ← best F1/AUC, reject rate 1.00

• The old 0.12 was genuinely too low; 0.75 is where a mis-scoped box becomes healthy (reproduces spike 3e).
• Item 3 and item 4 are complementary: a well-scoped (extent) box at 0.0 padding already beats a bare box at 0.75, but padding still adds separation on top, peaking at 0.5–0.75.
• 0.75 is the safer default — near-optimal when the box is well-scoped, clearly best when it isn't. No case for dialing back toward 0.12.

Caveats: LFM is a proxy for Cosmos; n=36 across 3 keyring instances; the tracker side (0.0) was not re-verified here (the probe has no tracked clips).

Tests

• New test_box_padding_is_conditioned_on_the_box_source asserts the grounded path widens (grounder padding) and the register-button path stays tight (tracker padding), across both enrollment paths.
• Full suite 175 passed, ruff and pyright clean.

🤖 Generated with Claude Code