fix(g4): emit measured floor-arena rest-z so libero_object pose_tolerance passes (RCA Regime A)

[KE7]

g4 task-object pose_tolerance drift — RCA + Regime A (floor arena) fix

RCA of the 66k publication run (run3) g4 task-object consistency drop
(41,515/66,430 = 62.49%). Full RCA: validation_run3/rca/g4_task_pose_drift.md.

Root cause (diagnosed, not band-aided)

The g4 pose_tolerance failures are NOT physical pose drift. xy is realized to
sub-millimetre and objects settle stably; 100% of the error is in z. The failing
quantity is the renderer's analytic, pre-physics spawn-z reference (obj.position[2]),
which diverges from the deterministic post-settle MuJoCo z.

This PR fixes Regime A — the libero_object (floor) arena, the largest single
deterministic regime (BEFORE: 0/420 = 0.00% pass on these tasks; max xy only 0.66mm,
all error in z up to 144mm). The libero_object suite maps to workspace class floor
(arena_surface_z(floor) = -0.115), and the kitchen per-class clearance does not
transfer — the suite authors several classes (ketchup, salad_dressing, milk,
orange_juice) in a different rest orientation, so the same class settles at a different
body-origin clearance. The clearances were also partly n=1 / missing.

Fix (Option 2: make renderer_z == settled_z, gate unchanged at 5mm)

floor is exactly the situation PER_ARENA_TABLE_CLASSES + measured |
clearances already handle for living_room/coffee — there were simply no *|floor rows and
floor was not threaded.

1. asset_metadata.PER_ARENA_TABLE_CLASSES: add floor so the renderer threads
   surface_class='floor' for floor-arena table-resters and object-axis variants.
2. scripts/measure_spawn_clearances.py: register the floor arena (10 basket tasks)
   with a position+object sweep (12 seeds), per-arena seed/subset configs (living-room/
   kitchen stay byte-identical), a determinism (spread) report, and an --arena= filter.
   Corrective-merge (>5mm only).
3. data/spawn_clearances_variants.json: +15 measured |floor rows (canonical +
   variants). All measured at 0.00mm rest-z spread (fully deterministic); merge is purely
   additive — zero existing rows removed or changed.
4. tests/test_spawn_z_consistency.py: 3 tests pin floor threading, presence of the
   measured |floor rows, and that resolved floor z lands in the suite band.

Determinism de-risk (STEP 0)

Settling each class via the production path across 8 seeds gives 0.00mm rest-z spread —
the 67mm cross-class "implied surface" spread in the RCA was a data-quality artifact of
the borrowed kitchen clearances, not physics wobble. No metastable residual on floor.

Verification

• Codegen: renderer now emits floor z matching the measured settle to <0.05mm
  (ketchup 0.1285 vs 0.1285 — was off 91mm; salad_dressing was off 144mm).
• Env before/after (scripts/verify_g4_floor.py, 10 floor tasks × position/object subsets):
  BEFORE 0/420 = 0.00% pass (max |dz| 144mm) → AFTER 560/560 = 100.00% pass
  (max |dz| 0.05mm, max xy 0.66mm unchanged, 0 g3/build fails).
• No regression: data merge is additive |floor-only; the code change only affects
  floor-workspace scenes, so the 191,285 currently-passing (non-floor) keys are
  untouched by construction. xy stays <1mm. Fixtures untouched.
• tests/test_spawn_z_consistency.py 96 passed; ruff + black clean on src/+tests/.

Scope note (refined taxonomy)

Re-analysis shows the original RCA "stacked (~10,439)" bucket is predominantly
table-resters in other arenas (table, kitchen_table, living_room_table, study_table) with
the same per-arena-clearance mechanism — NOT object-on-object stacking. True stacks are a
small tail (akita_black_bowl stacked_on {cookies, ramekin}). Those other-arena corrections
(incl. reference-arena canonical fixes and a few living_room_table metastable residuals to
flag) and the true-stack renderer offset are the same template applied per arena,
recommended as focused follow-ups to keep each change deterministic-gated and low-risk.
This PR lands the fully de-risked floor regime.

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com