Legacy count-selector plan scope repair — 2026-11-05
The concrete ReferenceError: transport is not defined is repaired in the
isolated callback regression in test/test-legacy-count-selector-webgl.js.
No production renderer or shader code changed. The emitted public COM fixture,
independent oracle, two-byte tolerance, historical module pins and accepted
positive evidence identity remain unchanged.
Evidence: scratch/runs/20261105-legacy-count-selector-plan-repair/.
The first executable test after the required releases and independent cleanup
was the unchanged count-selector --plan. It exited 1 at
capturedMatrixControls, on transport.batches.push in the extracted
port.postMessage wrapper. Tested main was
f8b9f49d4194505eaf3f311134e295ad4d5010dc; the relevant test/fixture and
backend/shader files matched published
ddc9e0e51136132e8c00725756ad91d45f6ad6ab exactly.
Actual faulty scope
The full ancestor browserFixture in
test/test-legacy-texture-rendered-contract.js declares transport,
queueBuffers, queueControl and heldPackets. Its real Worker callback
records batch headers before deciding whether delivery is held. The isolated
regression extracted that callback into a hand-built VM context which omitted
those locals. Non-strict init assignments could create accidental globals;
the first batch then failed on the missing transport binding. This was not a
renderer scope failure or a failing captured-matrix pixel control.
Git history explains the drift: f1dc763eb3 added the queue fields and
fbc8051625 added transport instrumentation after the isolated callback check
from bf257f136f. The repair executes the actual emitted declaration prefix
together with the actual emitted wrapper in a strict isolated realm.
It preserves the real fixture's transport instrumentation and Worker guards.
The regression executes init, immediate delivery, held batch plus following
message, and ordered drain. It checks captured sequence/index/command/header
metadata, shared control/buffer identity, no premature mutation, selected
endpoint buffer only, type-4 matrix only, adjacent byte guards and malformed
record rejection. In-memory omissions of each of the four bindings must fail
with the corresponding ReferenceError before delivery or matrix mutation.
The historical missing-identity helper mutation also rejects without changing
bytes. Parsing alone cannot satisfy these checks.
Related projected fixture failure
The focused test-legacy-projected-webgl.js --plan also exposed a pre-existing
unique fault seam ignored assertion. Its old ignored/early fault injectors
searched for stage.transformFlags = 0x103;; the current adapter conditionally
leaves type-6 FLOAT3/4 inputs unprojected. The test injectors now target the
current conditional assignment. Early division changes projected type-4/type-5
coordinates while preserving type 6. Nine isolated controls execute the actual
adapter transform block, checking transformed coordinates, divisor, geometric
RHW and stage flags for baseline/ignored/early variants. No production adapter
or sampling path was changed. The original related failure log is retained.
Validation and evidence limits
- Count-selector
--plan: passes all 155 states, compiler/Device checks and the strict isolated callback positive/negative controls. - Projected WebGL
--plan: passes, including nine executed adapter fault controls. Coordinate-contract and base public-texture plans pass. - Bounded public x86 COM software fixture: COUNT2/3/4, selector 7, scale 16, independent geometric W, both shipped modules, direct software and actual native Worker replay. Four routes, 12 successful public draws and 36 independently recomputed pixel observations pass. Both queued routes hold batches before live mutation, replay production commands and report zero queue fallback. These are software results, not WebGL pixels.
flock /tmp/wine-build.lock bash tools/build.sh: passes.
baseline-identities.json and unchanged-contract-final.json record source,
emitted-fixture and oracle hashes. No historical fixture pin was renewed.
The retained historical module prerequisites may therefore reject current
modules; this repair does not claim those historical hashes identify today's
build. The current tested modules are:
| Module | SHA-256 |
|---|---|
| Normal | 10462bd9693cf4f51d382d33767674b736b84e74346f328306d59e35c5e60031 |
| Compat | eb59d32227cc18233f93c7b757e0700f9558f0a31af7e3d38090669d4fc15158 |
The published quotient-gradient repair
already has 598 rendered observations across 20 routes. reused-598.json
records the retained result hash, counts and matching current renderer/module
identities. That matrix was reused, not rerun. The count-selector emitted
fixture SHA-256 remains
11dbf0bb5d7f43297d45ae522148cdf1f3e096c32c9ad7a7741049bfc2217ffd;
the independent coordinate oracle remains
f2968810d7ec9a961861d90328c4f47a95b94a1a3d8e8170fc1d04adf3e56025.
Plans and isolated transport checks are not pixel evidence. The revised fault
injectors have isolated execution coverage, not new rendered-fault coverage.
There is no new GL, physical-GPU, native-Windows, original-game or gameplay claim.
Ordered ReturnFire1435, Lords1450 and Strong1500 FINAL releases were observed before shared edits or executable tests. Independent cleanup verified 187 PIDs absent, 90 ports refused, 16 profiles absent, four locks free and no heavy runtime before reproduction. Git SHA plus relevant binary diff and the new note patch preserve recovery; no source archive, module copy or media copy was made. No browser, box, subworker, public deployment or platform rejection occurred. Box ID/state/expiry are N/A; lifecycle ownership stays with the coordinator. Own command processes and the board watcher are cleaned up before the final runtime/file release. Next: continue the broader advertised texture behavior audit; this plan repair requires no additional matrix run.