fix(g3-budget): align sweep→eval resolver + re-key resolver on expensive geometric axes + document/retry residual

[KE7]

Summary

Fixes the run3 g3 publication finding (g3 = 96.85%, 2,092 RejectionException fails). The budget triage (rca/g3_budget_triage.md) diagnosed these as three populations — and showed "just align the sweep to the eval budget" is insufficient:

Class	Count	%	Remediation
sweep-cap artifact	~1,304	62.3%	R1 align sweep budget → eval resolver
resolver-gap	~748	35.8%	R2 fix resolver preset-containment logic
genuine-hard	~40	1.9%	R3 document §6 + bounded resample retry

R1 — align the sweep budget to the eval resolver

The sweep sampled every condition at a flat --max-iter 2000 — tighter than the eval/gym path's resolve_iteration_budget() — producing false g3 failures for conditions the real pipeline samples fine. --max-iter now defaults to omitted → the sweep resolves the budget per (task, axis-subset) through the same resolver the eval path uses, so sweep and eval agree by construction. An explicit --max-iter N still overrides (back-compat / debugging). Routed through the resolver — no hardcoded flat bump.

R2 — fix resolver under-provisioning (the principled core fix)

resolve_iteration_budget gated the combined/full budget on full preset-superset containment — including the geometrically-free appearance axes (object/camera/lighting/background, ~1 iteration each). A subset carrying all the expensive geometric axes (position+robot+distractor) but omitting one cheap axis was starved at the 5000 floor despite being geometrically as hard as combined.

Re-keyed on EXPENSIVE_GEOMETRIC_AXES = {position, robot, distractor, articulation} — a preset's budget applies iff (preset_axes & EXPENSIVE) ⊆ request. Still monotone in axis-set inclusion. position,robot,distractor now resolves to 55000 (was 5000); cheap subsets stay at 5000 (no balloon). Logic fix, not a constant bump.

R3 — document + retry the residual

• (a) The robustly-over-constrained full-perturbation cells on the 5 worst multi-object tasks fail even at the 55000 eval ceiling across seeds — documented as §6 criterion-6 known-hard in data/known_hard_cells.json + docs/scenic_iteration_budget.md. Not masked with a larger budget.
• (b) Added a bounded resample-on-RejectionException retry (G3_RESAMPLE_RETRIES=2) to the sweep's G3 stage for the heavy-tailed high-variance tail (e.g. bowl_on_stove: FAIL seed 0, PASS 4,441 iters seed 1). Mirrors the eval reset() settle-loop resample policy — bounded, never unbounded, never a budget bump.

Before/after evidence (live, multi-seed)

Per the triage's own single-draw classification (one seeded draw at the 55000 ceiling classifies a cell against every budget threshold at once). Pass rates over seeds:

(task, subset)                                          new_budget   OLD@2000   OLDeval   ->  NEW    n_iters
-- resolver-gap (R2): OLD eval@5000 fails, NEW@55000 passes --
put_the_bowl_on_the_plate × position,robot,distractor      55000      0/5       0/5      ->  5/5   19629,50510,5108,11343,10892
put_the_bowl_on_the_plate × position,object,robot,distr.   55000      1/5       1/5      ->  5/5   37336,5111,15411,46158,758
open_the_top_drawer × position,robot,distractor            55000      1/5       1/5      ->  5/5   107,27985,36535,8303,6083
-- sweep-cap-artifact (R1): eval already sampled these; sweep@2000 was just capped too low --
put_the_bowl_on_the_plate × combined                       55000      0/5       5/5      ->  5/5   11157,12967,8130,6082,5281
KITCHEN_SCENE2_..._on_the_plate × combined                 55000      4/5       5/5      ->  5/5   2809,267,263,68,142
open_the_top_drawer × combined                             55000      1/5       4/5      ->  4/5   1500,12943,11463,8038,FAIL*
-- genuine-hard (R3a): still FAIL at the ceiling, documented as §6 known-hard --
KITCHEN_SCENE7_open_the_microwave × full                   55000      0/3       0/3      ->  0/3   FAIL,FAIL,FAIL
bowl_on_wooden_cabinet × full                              55000      0/3       0/3      ->  0/3   FAIL,FAIL,FAIL
-- cheap controls: budget stays 5000, still pass fast (no balloon) --
put_the_bowl_on_the_plate × object,camera,lighting          5000      2/2       2/2      ->  2/2   1,1
alphabet_soup × position                                    5000      2/2       2/2      ->  2/2   2,1

* the single open_the_top_drawer × combined FAIL@55k is the high-variance straddler the R3b bounded resample retry targets.

Expected post-fix g3 publication rate: ~99.94% (R1+R2 resolve ~98.1% outright) + the documented ~40-cell §6 residual reported honestly rather than as silent failures.

Tests / CI

• R2 regression (test_scenic.py): resolver-gap subset inherits combined budget; cheap subsets don't balloon; expensive keying stays monotone.
• R1 + R3b regression (test_sweep_g3_budget.py, added to CI Tier-1): G3 budget routes through the resolver; explicit --max-iter overrides; cheap subset no balloon; bounded resample-then-pass and resample-is-bounded(→honest fail).
• Tier-1 (219 passed) + exact CI validation-smoke (6/6 G0–G3) green locally; ruff check, ruff format --check, black --check all clean.

🤖 Generated with Claude Code