feat(g4): accept valid alternate physical rest in pose_tolerance (SEMANTICS change — USER sign-off)

[KE7]

Summary — g4 pose_tolerance alternate-rest acceptance (net-add scoring change)

Closes the living_room metastable §6 pose_tolerance residual via the RCA-surfaced
Option §3c (a scoring change), NOT a data/gate/resampler hack. This changes what g4
certifies for the paper, so it is intentionally scoped for USER sign-off before merge —
EA can verify mechanically but must not merge on verification alone.

Chosen semantics (definition b — physical-rest validity), and why

The strict gate pos_err ≤ 5 mm AND rot_err ≤ 1° demands the settled pose be an exact fixed
point of the emitted pose. RCA (g4_fixed_point_settle.md §3, g4_metastable_residual.md §6)
showed a deterministic tail of tall/multi-face objects settling into a genuine second energy
minimum — upright, on-support, in-region, but a few mm past the 5 mm gate; iterating the
settle does not relax them. "Placed correctly" is a statement about the rest configuration,
not about the settle being a numerical fixed point of the emitted xy.

A settled object now passes iff EITHER:

• STRICT — pos_err ≤ 5 mm AND rot_err ≤ 1° (unchanged); OR
• ALTERNATE REST — all of: at rest (net drift of the vibration-averaged settle-tail
  position below tol), upright (env-settled vs canonical ≤ 1°), on its support
  (|Δz| ≤ ¼·object height), in its region (xy drift ≤ object planar half-extent, capped
  5 cm). Class checked separately; the alt-rest path scores the same body_id.

Acceptance is STRICT ∪ ALT-REST → a strict superset → net-add by construction. Every
threshold is anchored to the object's own measured geometry (asset-registry dims) or the
existing rotation tol — none tuned to the tail.

Rejects (never a mask), each with an explicit reason: fall-through/wrong-support
(off_support), out-of-region (out_of_region), tipped (tipped), moving (not_converged),
missing data (conservative decline → strict only).

OLD-vs-NEW A/B — scripts/ab_g4_altrest.py, 1360 objects, 4 arenas, position+object, seeds 10

metric	value
OLD (strict) pass	1348 / 1360 (99.12%)
NEW (strict ∪ alt-rest) pass	1352 / 1360 (99.41%)
FAIL→PASS (closed tail)	4 — all living_room / butter (69/73 → 73/73)
PASS→FAIL	0 (guaranteed by STRICT OR ALT; empirically confirmed over 1360)

• Closed tail: butter (multi-face stick) settling on a secondary flat face ~7 mm lower —
  pos_err 6.9–7.4 mm, all on z (dz ≤ 10 mm ¼-height band), xy ≤ 1 mm, upright 0.001–0.48°,
  converged (2.3 mm / 0°). A textbook valid alternate physical rest.
• 8 residual fails all tipped (settled orientation > 1° from canonical — a slide-with-tilt,
  conservatively left flagged). Raising upright_tol_deg would close more at the cost of
  redefining "upright"; kept at 1° = the strict definition (maximally defensible default).
• Injected-bad rejection (per object, 1360×): fall-through / out-of-region / tipped /
  not-converged all rejected (1360/1360 each); converged-valid control accepted (1360/1360).

Implementation

• consistency.py — assert_pose_tolerance reports strict_pass + alt-rest verdict/reason;
  accept_alt_rest=False recovers the exact legacy gate (used by the A/B OLD arm).
• simulator.py — captures a non-contaminating settle-tail convergence signal (net drift of
  the vibration-averaged body position over the last 10 settle steps; a live qvel read is vacuous —
  velocities are zeroed after settle). Scored pose byte-identical to before.
• gym_env.get_object_state — exposes settle_conv_lin/ang + canonical_orientation (the
  Scenic object's own orientation does not coerce for real objects, so upright is measured
  env-settled-vs-canonical).

Guards honored

• NET-ADD only — 0 pass→fail over 1360 objects. No resampler change, no data forcing,
  no strict-gate widening. xy bounded to the object footprint. fixtures = 7.

Verification (green from first run)

• Tier-1 pytest 194/194 (incl. 30 consistency tests: closed-tail + explicit negatives —
  fall-through, out-of-region, tipped, moving, missing-signal, canonical-upright path).
• validation-smoke G0–G3 6/6 pass. ruff + ruff-format + black clean on src/ tests/.

New g4 definition (crisp, for USER sign-off)

A settled object passes g4 pose_tolerance if it matches the emitted pose within 5 mm / 1°
(unchanged), or if it is a valid alternate physical rest — at rest, upright (≤ 1° from
canonical), on its intended support (Δz ≤ ¼ of its own height), and within its own footprint of
the sampled xy — while rejecting fall-through, out-of-region, tipped, wrong-object, and
still-moving settles with explicit reasons. It changes what g4 certifies from "the settle is an
exact fixed point of the emitted pose" to "the object came to rest, correctly placed on its
support, in its region, upright."

Full write-up: docs/g4-alt-rest-semantics.md.

🤖 Generated with Claude Code