fix(scenic): SAT-correct AABB clearance for fixture pairs — closes c.1

[KE7]

Root cause (campaign caveat c.1, drawer-family G3 residual)

The require clauses for object↔fixture and distractor↔fixture clearance in src/libero_infinity/renderer/scenic_renderer.py used a radial-diagonal form:

require (distance from obj to fixture) > (hypot(w_f, l_f) + hypot(w_o, l_o)) / 2

By the separating-axis theorem, two axis-aligned AABBs (which is what every LIBERO fixture / object is) are non-overlapping iff:

|dx| > (w_a + w_b)/2  OR  |dy| > (l_a + l_b)/2

The radial form is ~√2× more conservative than necessary in the symmetric case and worse for elongated fixtures (flat_stove 0.36×0.20, desk_caddy 0.14×0.42, wooden_cabinet with door 0.30×0.30). For wooden_cabinet vs a 0.16×0.16 bowl the radial threshold is 0.325 m vs the AABB-correct 0.23 m on either axis — a sizeable spurious shrinkage of feasible region.

Combined with SAFE_REGION (0.70 × 0.50 m) and 4 sampled distractors, plus the existing distractor-pair and distractor-object clearance conjunctions, the result is the exact cardinality-monotone rejection signature documented in rca/stage4_c1_addendum_10_task_footprint.md:

cardinality	G3 fail rate
2	0.064%
5	0.242%
8	0.769%

(All 50/50 worst-conditions in reports/worst_50_rca.md are c.1 RejectionException at scenarios.py:437.)

The original comment in the code justifying the radial form ("OR form permits diagonal corner penetration") is mathematically incorrect: disjoint x-projections imply disjoint AABBs regardless of y, and vice versa — that's the entire content of the separating-axis theorem.

Fix

• Replace distance > _footprint_clearance_xy(...) with the SAT-correct per-axis OR-form for object↔fixture and distractor↔fixture clearance, exactly matching the form already used for object↔object pairs at line 866-867.
• Add _footprint_clearance_aabb(dims_a, dims_b) -> (dx_min, dy_min) helper.
• Retain _footprint_clearance_xy for non-renderer callers that need a single scalar clearance (e.g. settle-drift checks); update its docstring to clarify that the renderer uses the SAT-correct form.

Why this is principled, not a band-aid

• This is not loosening a physically-required clearance — the radial form was over-conservative for AABB geometry. The OR form is the tightest possible non-overlap condition. No real overlap becomes admissible.
• No magic numbers were tuned (no min_clearance lowered, no maxIterations bumped, no task-specific bypass added).
• The fix is uniform: same form everywhere in the renderer (object↔object already used it).

Smoke evidence (scripts/smoke_c1_drawer_family.py)

10 c.1 tasks × 5 axis-subsets (cardinality 2 → 7) × 5 seeds = 250 conditions.

==============================================================================
Total conditions:   250
Passed:             250  (100.000%)
Failed:             0
Max iters observed: 2553 (cap=25000)
==============================================================================

≈10% of the 25k budget is the worst observed; previously these tasks were budget-exhausting at 25k. The full output is reproducible via:

.venv/bin/python scripts/smoke_c1_drawer_family.py

Test plan

• pytest tests/test_renderer.py tests/test_perturbation_audit.py tests/test_settle_correctness.py → 58 passed
• python scripts/smoke_c1_drawer_family.py → 250/250 pass at 25k iter cap

References

• ~/.omar/ea/4/validation_run/reports/worst_50_rca.md
• ~/.omar/ea/4/validation_run/rca/stage4_c1_addendum_10_task_footprint.md
• ~/.omar/ea/4/validation_run/rca/stage3_open_drawer_persistent_25k.md

🤖 Generated with Claude Code

[KE7]

CI audit (#17) — RCA for the red Tier-1 unit job here

The two failing tests in tests/test_scenic.py:

• test_position_mode_adds_fixed_fixture_clearance_constraints asserts "require (distance from " in the emitted code.
• test_distractor_mode_compiles asserts "(_n_distractors <= 0) or ((distance from distractor_0 to wooden_cabinet_1)" in the emitted code.

This PR's whole point is to swap the (incorrect, circular) distance from X to Y > r formulation for the SAT-correct AABB clearance abs(X.x - Y.x) > rx or abs(X.y - Y.y) > ry. The generator is now emitting exactly that — you can see it in the run-log diff. The tests, however, still match the old circle form.

Action required on this branch: update those two assertions (and any sibling assertions in test_scenic.py) to match the new AABB form, e.g.

assert "abs(" in code and ".position.x -" in code and ".position.y -" in code

or more strictly:

assert any(
    f"abs({obj}.position.x - {fix}.position.x) >" in code
    for fix in fixtures
)

Once the tests are updated to assert the new formulation, this PR's Tier-1 will clear. CI infrastructure is fine — fix belongs here, not in the CI branch.

ci/audit-and-stabilize (PR #17) does not attempt to patch around this — per the no-band-aid rule.