Original LBreakout initialization cost and board readiness

Evidence: scratch/runs/20261110-lbreakout-init-cost-cause-0142/ (ignored). Initial repair shipped as 5cbb2a8f4 on main before the sole original run. Accepted fixture 4fc62b7d7 was independently hash-checked and reused unchanged.

Cause and repair

The previous two-second observation entered original local-init after 1281ms, leaving 719ms. Its one-address trace still globally disabled transfer chaining and the micro-op tier. Filtering callbacks does not remove that execution cost. The generated level-start observer now uses the existing compiled breakpoint and target-aware chain exit. It follows original entry/return boundaries without guest writes, artificial handler calls, module edits, or provider changes. Moving the target retires the old breakpoint's skip-once block first. The unarmed run and metric functions retain native function identity.

Completion polling also repeatedly fetched and hashed the entire 2,186,068-byte Wasm through CDP. The consumer now validates bytes before arming and checks the immutable owning instance, endpoint, command, and DLL binding on each poll. Full byte validation still brackets the first capture. This removes downloads from the unchanged two-second completion/presentation budget.

Offline evidence

tools/lbreakout-init-cost.test.cjs executes actual original EXE, SDL and CRT instructions through the actual compiled module. Alternating matched samples cover original background tiling/software blits and all 25 original levels parsed through original fopen/fgets. Identical complete pixel outputs, 125 precise hits, four moving boundaries, debugger conflicts, native restoration, and the full generated preload all pass.

Five background-work pairs (25 calls/sample) had median milliseconds: disabled 58.38, trace 62.01, target 55.96. Three complete-levelset pairs: disabled 11.62, trace 14.38, target 11.04. Trace enabled the global tier guard even with zero callbacks; disabled/target left it zero. These isolated workloads demonstrate instrumentation cost without claiming to reproduce every browser service or all initialization. Retained logs cannot exclude earlier Worker RPC stalls; no completed HTTP request followed the earlier arm.

Actual generated-consumer controls exercise success, wrong command/module binding, incomplete work, failure, presentation refusal, and foreign immutable instance. Negatives withhold the first image and prohibit subsequent input, capture, and quit. The previous consumer made three full-module transfers during completion; the initial repair made one, and the final offline repair makes zero. Byte validation before arming and around capture remains enforced. Ten inspection tests include native unarmed identity and emitted-function presentation behavior. Level-start, pause, and original SDL input tests pass; locked builds pass.

Sole live result and terminal boundary

Every dependent image was individually model-reviewed. Native original menu navigation reached the real Late At Night board, level 1/25, Michael score 0, bricks, paddle, attached ball, and lives visible. From Enter arm: local-init entry +505ms, return +829ms (324ms inside), game-run +830ms, fade return and first display completed +1210ms. The board image passed the presentation gate.

The next ordinary P command reached the original pause handler, original pause display return, state 20, handler return, and released P. Guest completion took 1921ms. The remaining full-module owner revalidation exhausted the presentation budget and failed at the original deadline before any pause image. Immediate owning cancellation followed. There was no later protocol, input, capture, retry, quit, extension, or second box. Visible pause, score progress and clear are unconfirmed. This is not a first-level-clear acceptance.

That terminal evidence caused the final offline removal of the redundant pre-presentation module download. Capture already validates the full bytes before returning an image. The final change is tested offline only; this task's one-live-attempt authorization is exhausted. Pause's remaining trace cost may still constrain the two-second bound; no new live success is claimed.

Lifecycle and next step

Only new box bx_chfab9w3, immutable UUID 42b0055d-0daa-4cdb-8d21-182048b42b2c, was touched. The independent watchdog was armed before allocation with deadline 2026-11-10T02:20:26.691281Z; initial API expiry was 2026-11-10T02:20:26.467Z. All 115 artifact references (100 unique) were recovered and independently reread for SHA/size; seven PNGs passed CRC/decompression, and no gzip artifact was present. Actual launcher wait and separate stdout/stderr EOF were verified before normal stop.

Exactly one delegated unforced stop, stop_fb0a2cdd6cf8f5ac, completed at 2026-11-10T02:02:12.778Z. Final state is archived, expiry null. The watchdog was actually waited and reaped with exit 0 and separate EOF. Exact deadline, retrieval, final API state and operation receipts are retained in the run.

Next work requires separate authorization for a fresh bounded original attempt to review visible pause and then launch/score. No live continuation remains authorized in this worker. Shared providers, Q3 and Nox were read-only.