MCM Worker observation: uninterrupted slice retains synchronous render waits

The bounded investigation closes with a measured blocker and no production change. Fresh CDP identity reads on the actual production dx_boids Worker route waited for the enclosing guest slice to return. Render completion notifications and Atomics timeout returns inside that slice did not service those reads. This is a local Chrome/SwiftShader observation prerequisite, not an MCM performance result or an original-game acceptance result.

Evidence is in scratch/runs/20261107-mcm-worker-cdp-latency/. The interrupted launcher2266488/child2266491 session was not reused. Its retained trace and samples were recovered; no additional runtime was needed. Source commit is afe89042a8956b7fa0278bf561ea8df95ea853bf; source-sha.txt, initial-relevant.patch and new-wininit-boot.patch preserve relevant source identity. The executing module recovered through CDP is SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. HTTP body hashes and lengths, including the existing media paths, are retained in each sample's served.jsonl; no source or media archive was made.

Measured boundary

Run python3 scratch/runs/20261107-mcm-worker-cdp-latency/verify-boundary.py. It checks the actual guest identity-expression hash, zero guest recorder drops, enclosing slice, intermediate completed waits, and reply alignment:

Slice Requested blocks Identity latency ms Reply after slice end ms
28 100000 2479.71 1
30 100000 3438.30 2
32 100000 4049.98 2
34 100000 1693.63 2
36 47000 2039.51 4

Every row crosses an earlier completed _waitFor without receiving its reply. For slice28, the read starts at 14:54:08.899Z; a completion wait returns at 14:54:10.218Z, then the next wait includes a 1000-ms Atomics timeout returning at 14:54:11.219Z. The final completion wait returns at 14:54:11.375Z, the slice ends at 14:54:11.378Z, and the CDP reply arrives at 14:54:11.379Z. Thus neither an intermediate completion nor the existing timeout is an observation yield. Integer wall-clock timestamps have millisecond precision; these gaps include transport and do not establish submillisecond ordering.

For that fully covered delayed read, analysis.json measures 2475 ms of overlap with guest _waitFor and 2474 ms with nested Atomics.wait. Renderer overlap includes 1205 ms in getError and 1153 ms in fenceFlip. These are nested categories, not additive CPU costs. The guest CPU profile attributes 27283 of 27462 sampled wall milliseconds to _waitFor; renderer getError accounts for 10618 of 27471 ms. Renderer fence polling can yield its own task while the guest remains synchronously blocked waiting for replay completion.

The renderer recorder reached its 30000-row cap: 41892 rows were dropped, and its last retained timestamp is 14:54:13.306Z. Later guest delays remain fully recorded, but their complete renderer breakdown is unknown. The first guest installation request also waited 3172 ms before instrumentation existed; it must not be treated as a recorded guest slice. Two failed span installers are retained separately and closed without input.

The observed 1000-block slices took 0.145–0.980 ms; the recorded 47000–100000-block slices took 1909–4051 ms. This does not establish that a smaller budget fixes latency: a single synchronous render import can retain its wait regardless of budget. lib/guest-worker.js calls ex.run inside an uninterrupted handler; lib/d3d-command-stream.js loops synchronously in _waitFor. The page's adaptive budget is updated only after sliceDone and only for qualifying budget-consuming slices. No budget experiment was made.

Executable next step and correction constraint

The next read-only probe is prepared and its wrapper transparency controls pass (no new runtime was launched during recovery):

python3 scratch/runs/20261107-mcm-worker-cdp-latency/prepare-next-probe.py
CASE=boundary-next SPANS=1 BROWSER_DEADLINE=$(($(date +%s%3N)+180000)) \
  node scratch/runs/20261107-mcm-worker-cdp-latency/next-browser.cjs

Use a fresh CASE directory each time. This captures twice through the unchanged live identity chain and closes its own browser. The prepared run-only recorder retains all guest spans, renderer calls taking at least 2 ms and errors, plus aggregate counters for every completed renderer call. Its 30000-row cap and dropped count remain explicit. It avoids spending the cap on fast fence polls; accept full coverage only if its dropped count is zero. Correlate one guest identity request with complete renderer service and GPU fence completion. The remaining unknown is how much of renderer service is avoidable native16 error-query synchronization versus required GPU work; removing getError is not justified by this profile alone.

A correction must either reduce proven unnecessary renderer synchronization while preserving error handling and native16 draw order, or implement an explicit resumable guest import boundary preserving completion tokens, render buffers, scratch ownership and guest continuation. The current synchronous Wasm import cannot safely become a Promise by changing its wait loop. Any such production change needs a failing-before/pass-after regression and locked build. No speculative timeout, yield, step-budget or freshness change is proposed.

Safety and cleanup

The full live owner/module checks, 12000-ms freshness, exact numbered ACK, independent UP and sticky UNKNOWN remain unchanged. tools/mcm-bound-input.cjs remains SHA-256 9df1455f51f91d76e17890ff4d90afe9264759d3e30e224a9a7569e29b1f8062. The retained samples only captured observations: no input dispatch, original guest, boat, accepted fixture matrix, held-lane retry, or guest/config/clock edit. No Dungeon Keeper helper/test was touched.

All four retained browsers exited code0. Recovery verifies their exact driver and Chrome PIDs absent, server/debug ports closed, profiles and writer locks absent. The predecessor's orphan messageboard watcher (PID2275380, exact retained PID and command verified) was terminated with SIGTERM. No browser required termination. No test box was assigned: ID/state/ expiry are not applicable, and no box lifecycle action was performed. recovery-cleanup.json, boundary-verification.json and the final evidence manifest record those checks. The predecessor locked build and run-only probe controls passed; this recovery reruns the recorder controls and boundary checks. Only this documentation is committed. Free space was sampled above 2 GB.