Fresh target clear-value cache repair

Starting revision: 912cd44a9e0b4ffedf106ecc80031351773f5302, retaining 77180c76b312efe465ad91e60286d40330d77d0d and producer ffe822a9d50ad79439330609915eb84475acd2bc. All required commit identities and the rejection's 12 evidence seals were verified. Strict trusted SSH reconciled remote main at f6571101e38985b3fa34d328f2ccadd9735a7573 before checkout. This task publishes a repair branch for independent review; it does not integrate main.

Behavior and owning boundary

The exact independent fresh-target probe failed before editing: a preceding wrapper clear cached depth .25/stencil 52, but real named target initialization wrote depth 1/stencil 0 directly to GL. The subsequent requested .25/52 clear was suppressed by the stale cache and used 1/0. Existing-target and cold-cache controls both passed on that unchanged rejection revision.

WebGLBackend.bindColorResource now initializes depth/stencil through the existing clear(null, mask, 1, 0) wrapper. The real initialization, mask, allocation, attachment and fresh-store lifecycle remain intact. The wrapper updates the shared clear-value cache when it changes GL state. Subsequent requested values can then be written or safely suppressed against accurate state.

A narrow audit found the same ownership violation in bindRenderTarget: it wrote stencil zero directly, then called cached clear without a stencil value. The new isolated implicit-target test reproduces its cache/GL mismatch on the reverted backend. That initializer now passes depth 1/stencil 0 through the same wrapper, retaining color and fresh depth/stencil initialization. Only the wrapper now contains direct clearDepth/clearStencil/clearColor calls in this file. No adapter, descriptor, existing probe, oracle or initialization was removed.

Validation

Evidence is under evidence/d3d9-fresh-target-cache-repair-20261024/. tests.json records exact argument arrays, exits, elapsed time and fresh memory checks; execution-protocol.json records the common command prefix and limits. Each Node command used timeout 60s node --max-old-space-size=256, CPU soft/hard limits of 60 seconds and fresh MemAvailable >= 805306368 bytes.

• Unchanged fresh-target probe: rejection exit 1, repaired exit 0. Existing-target/cold-cache controls: 0 before and after. Its SHA256 remains 4b550a8b7966228cddb3957c789803835c1df05bc128c45c5fac14a3b7e86408.
• Loading only the exact rejection's GPU backend into an isolated process with node -r /tmp/d3d9-fresh-target-cache-repair-oct24/revert-backend.cjs makes that unchanged probe exit 1 for its original named-target assertion.
• New test/test-d3d9-clear-fresh-target-cache.js: 0 for eight named cases covering D16 and D24S8 depth-only, stencil-only, combined clears, requested default/nondefault values, warmed cache, repeat suppression, fresh second attachments, existing-target return and implicit/named switches. It exercises actual Device/binding/depth-storage/cache/clear methods. It checks cache/GL agreement, initialized untouched depth/stencil, color preservation, unrelated viewport/depth-range/blend state and retained first-use initialization. GL and color/texture allocation are mocked. --revert exits 1 at requested named depth; --implicit-only --revert independently exits 1 at implicit stencil cache agreement. Final candidate rerun passes after adding preservation assertions; both mutations select the exact rejection backend through Module loading.
• Original independent normalized-depth test and --probe: 0 unchanged. Its exact-main --probe --baseline exits 1 for the original normalized-depth reason. Existing focused depth-cache suite: 0; its exact producer adapter --mutation: 1 for the first normalized-depth assertion.
• Producer clear-contract suite and its seven semantic rejecting controls: 0. Compiler path, five candidate Git blobs/full-file hashes, ordered balanced extraction ranges and byte-exact 5257-byte source were revalidated first. Installed compiler was executed in place. This is only a five-function, one-64KiB-page excerpt producing 755 module bytes, never a full module build.
• Seven cheap existing suites pass: GPU wrapper, shipped browser dependency ordering, neutral command queue, shared render worker, software synchronous present, D3D9 shared worker and present cadence. Changed-file syntax and git diff --check pass. Original four test files remain byte-identical.

These observations establish source behavior at a mocked GL boundary. They do not establish real GL driver pixels, complete module compilation, rasterized draws, browser execution, guest compatibility or gameplay.

Resource, publication and recovery

New explicit non-cone sparse checkout: /tmp/d3d9-fresh-target-cache-repair-oct24/repo, branch task/d3d9-fresh-target-cache-repair-oct24. Required paths were the supplied donor manifest plus exact rejection changes. Read-only object donors and installed Node/compiler were used in place; only this checkout's alternate list was flattened. Deep alternate warnings were retained, not treated as missing objects. Root tracked files/index/refs and all donor checkouts stayed read-only. Coordination used append-only echo >> messageboard.txt followed by tail.

Input manifest measured 3534248 bytes. Original projected peak 56388608 bytes remains charged against 67108864, including checkout/Git, generated/temporary evidence, complete external session/events/logs, positive SQLite growth, future outputs and an 8388608-byte reserve. Per-file limit is 25165824 bytes; post-projection host free floor is 2982150144 bytes. Task-root admission, standing-tools, resource-final and cleanup receipts preserve measurements and history; capture accounting retains early slice overages and prior allowance. Fixture/path and initial metadata-recording corrections are disclosed in execution-protocol.json; failed raw logs were retained. No source oracle was changed to accommodate a failure.

source.diff, regression.diff, probe hashes, prior seals, raw logs and artifact seals allow independent review. Task-root publication/committed-verification receipts record the actual commit and independent strict-SSH remote SHA check. To recover, fetch this task branch with the existing trusted identity/known-hosts configuration; do not alter trust settings. Retained runner recipes assume they are placed back in the task root (parent of repo); compiler donor identity is pinned in the checkpoint. No generated binary requires recovery.

No box was assigned: ID/state/expiry N/A; lifecycle remains coordinator-owned. No runtime, guest mount argv, worker launch, panes, deployment or main merge. Own board watcher is terminated at release; evidence remains for coordinator review. This worker returns ownership and stops after verified publication.