0.2 — smoke_tram was ONE assertion (`!!getObjectByName('tram')`), a printed SKIP + early return
on absence, and it PASSED ON THE FENCED PATH because tram.js:110-114 adds a named EMPTY group.
Rebuilt STRICT, four arms + both controls demonstrated every run: RUNS (route/stops/metres, and
the body measurably moves under the SHELL'S own rAF loop) · BEHAVES (an isolated tram on a
throwaway scene, fixed dt: traverses, dwell time accounts to exactly dwells x 3.5 s, max cruising
tick <= SPEED*dt, every discontinuity bounded by 2*LANE with >=1 end reversal, and it rides
3.2 m off Lane A's OWN main-edge centreline) · GENUINELY ABSENT on the fenced katoomba_real
(fenced verdict, 0 stops, 0 group children, 0 m moved) · OFF under ?tram=0 and ?classic=1.
CONTROL A points the run-predicate at the fenced town and CONTROL B points the absence-predicate
at the running tram; both must go red or the gate fails. A missing subject now FAILS.
0.5 — classic_regression made no network assertion at all for 21 rounds, which is why R36's
fetch-surface widening needed a brand-new smoke. It now watches the wire from before the first
byte, with the allow-list DERIVED FROM THE v2.0 TAG at gate time (git ls-tree + the v2.0
manifest), never a deny-list. The four v2-era depot street GLBs and the five record-shop fittings
are BASELINE and admitted by provenance, not by exception. Same-origin is NOT a free pass -- the
R36 breach WAS same-origin, so /models/peds is enumerated by the v2.0 tree (exactly the
covenanted 21 files). Anti-vacuous: a required v2 core must be observed or the clean verdict is
refused. CONTROL, run every time: injecting the R36 breach itself (models/peds/woman_dj_01.glb)
plus a never-existed asset turns the clean verdict into exactly 2 violations.
Measured and carried, not fixed here: booting the actual v2.0 tree and diffing its fetch surface
against today's ?classic=1 shows a POST-v2.0 delta -- 5 audio files (R11 f7dd44f) and
assets/towns/index.json (R20/R21). Allowed on the list so the suite is not permanently yellow
over 26 rounds of shipped behaviour, printed every run so it cannot go invisible. Fable's to
ratify or retire.
Also: tools/qa/budget_walk.py (R37 0.3, the measurement rig -- continuous-motion walk through the
real rAF loop, bearing sweep, dwell, and a stepwalk that reproduces the audit's method) and
`flags_check.py --only` for single-gate development runs, which prints a PARTIAL banner so a
subset can never be read as a suite verdict.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
23968856dc