feat(fixtures): seat alphabet_soup on flat_stove burner plate (un-exclude from grate)

[KE7]

Goal

Actually seat alphabet_soup stably on the flat_stove burner instead of the PR #32 exclusion (which routed it to the table via unstable_distractors). Prefer seating over exclusion (as done for the wine_rack / shelf cradles).

RCA — why the can tipped (and why the "grate" was a red herring)

• The flat_stove burner "grate" is a VISUAL-ONLY mesh (burnerplate, contype=0/conaffinity=0 — non-colliding). The real collision is a flat plate (collision top z=0.9305 world) + a base box, plus the stove KNOB raised to z≈0.955 off to one side (world x ∈ [-0.28, -0.155]).
• The burner flat plate sits +0.15 m in x from the fixture body; the knob occupies the body-centred region. The renderer centres distractor placement on the fixture body → the placement region straddles the knob and the burner-left edge.
• The tall alphabet_soup can (r≈0.029 / h≈0.075) lands partly over the knob, tips off it and tumbles off the stove entirely (measured long-settle drift 846–4635 mm; 3/7 region points fail the 50-step gate). Shorter classes survive the 50-step gate by luck. So it tips because of placement over the knob, not grate bars.

Fix — seat, don't exclude

The burner plate is a clean flat surface. Add a per-(class, fixture) distractor SEAT that relocates alphabet_soup onto the burner-plate centre at the MEASURED flat-plate rest:

• dx=0.15, dy=0.0, dz=0.1489, x_half=0.05, y_half=0.05, upright (default x90 — no tilt).
• dz=0.1489 is the measured flat-plate rest and supersedes the stale knob-contaminated variant clearance (0.1691).
• distractor_fixture_seat() / is_distractor_fixture_seated() in asset_metadata; spawn_clearance returns the seat dz first for a seated (class, fixture) distractor (renderer + simulator resolve the same z, lockstep).
• Renderer: a flat-top fixture slot with ≥1 seated pool class emits per-class correlated (xc, yc, xh, yh); the seated class lands at the seat region, every other class is distribution-identical (c + h*Range(-1,1) ≡ Uniform(c-h, c+h)).
• Data: flat_stove gains distractor_seats{alphabet_soup}, drops unstable_distractors. footprint / offset / height / top_z BYTE-IDENTICAL; 7 fixtures preserved.

Evidence

Deterministic on-burner seat (primary proof): placed via the production renderer→Scenic→sim pipeline, alphabet_soup now seats on the burner (on_burner=True), injected==settled to 0.12–0.24 mm at the 50-step production settle and 0.25 mm at a >1500-step settle, 0.00 mm inject-at-rest — reproducible across 3 tasks × seeds × a ±0.05 box (27+ trials).

Deterministic before/after (renderer's actual placement region):

• BEFORE (body/knob-centred): 3/7 region points fail @50-step; long-settle drift 846–4635 mm (tumbles off the stove).
• AFTER (burner-plate seat): all points 0.26 mm @50-step / 0.25 mm long; 0 fails.

Aggregate smoke (smoke_clearance_fix.py --seeds 8, fresh main 4fe683e vs branch):

• distractor: MAIN 182/194 = 93.8% vs BRANCH 192/209 = 91.9% (−1.9 pp). Branch places + passes MORE distractors in absolute terms (192 > 182); the rate dip is the scene-resampling confound (alphabet_soup now enters the flat_stove pool → RNG shift → +15 more distractor placements).
• task objects: 87.6% → 87.3% (flat).
• Matched-scene comparison (886 identical keys): overall pass IDENTICAL — main 788/886 vs branch 788/886 = 88.9%. The 16 pass/fail flips are symmetric borderline jitter (8 each way, 3–8 mm), net zero, none involving alphabet_soup-on-flat_stove. → no quality regression.

Scope / safety

• Distractor-only. flat_stove top_z / footprint / offset / height byte-identical (only the distractor_seats key added). Task seating untouched. Other fixtures untouched. place_half for other flat_stove distractors is byte-identical (cookies, not alphabet_soup, is the worst-case driver).
• Local CI green: ruff check + ruff format + black clean; Tier-1 (6-file) 174 passed; validation-smoke G0–G3 6/6 pass.

🤖 Generated with Claude Code