Hexen II early page profiling, 2026-11-20
The historical 234.525 ms task remains FUNCTION_UNKNOWN. This task retained a fresh page CPU profile and named call stacks, but neither establishes its causal function. No production repair or class/difficulty/level progress is claimed. Evidence and the exact function handoff are in scratch/runs/20261120-hexen2-early-profiler-capture-pipeline-original/.
The run-local pipeline starts profiling immediately after discovering the page target, before navigation, Worker source inspection, whole-host capture, or native bind. The corrected sixteen-path recipe/test manifest, physical tool paths, frozen bridge, and matching ninety-three prior controls were verified before the browser. No calibration was repeated. The CPU sampling interval was 1 ms, with a finite 180-second scope and a 48 MiB raw trace bound in assigned tmpfs. Start commands took 186.590 ms. Existing method wrappers, retained snapshots, CPU sampling and timeline instrumentation perturb the run; their CPU overhead is unknown.
The original served frozen bridge 91069c0446876b6528c72db09517e45a1fe3a7122a1444bc3dc5408f8ed22c2b, both actual Workers were verified as d3c95bc28f49bd5e1bd05e6c7536a240ebdfffd803a98c6f3d0372afec3f36d2, and all eighteen media references matched. Shared W3 bridge/host/renderer files were read-only. The whole 800x600 host and focused Single Player menu were personally viewed. Trusted host focus, native bind and manual-context capture checks passed. The reviewed Enter used a 90 ms hold. Physical UP at 17:18:32.417 UTC preceded native USER return, and the page status was UNKNOWN at 17:18:32.429 UTC (received at .433); parent closure followed at .435. The scope was permanently closed. No later application, input or frame requests, subsequent captures, or replay occurred.
The separate profiling-only cleanup sent Profiler.stop, Tracing.end, and Profiler.disable. It retained original/page.cpuprofile.gz (82,973 samples, 1,511 nodes; compressed SHA-256 de5c5b69e7415940e5b1fcc12344dedcf52ac91d8e787da71274e2981a3d8a90). Cleanup took 5,467 ms and is not a page-task duration. Timeline data and tracingComplete were not received. Raw CPU time deltas include 7,992 negative entries; the analysis sorts reconstructed timestamps and uses nonnegative adjacent gaps, rather than assigning signed costs to functions. Occupancy estimates are not measured function CPU time.
Named sampled stacks include step → runMainThenEnd → runMain → slice → slice → _ask → Promise executor at served lib/guest-thread-host.js:619/639, and prepareEquivalent → verifyPageEquivalence at augmented host.js:8376/8424/8297. These are candidates for inspection, not causal findings. The latter belongs to socket preparation; the former is cumulative guest-command transport work. Dominant (program) and (idle) samples do not identify a JavaScript function. The profile interval and command wall interval disagree by 647 ms, exceeding the 26/132 ms start/stop command bounds. The saved affine projection is explicitly provisional and cannot establish capture-specific attribution. Fresh manual captures took 97/188/160 ms; these durations do not attribute the historical overlapping 170 ms source/capture and 157 ms socket/cancel intervals.
The precise handoff is function-handoff.json, with source identity, candidate call stacks, missing correlation, and unchanged 250/500 ms guards. Before another separately authorized original, verify profiling-only trace completion and clock correlation on a non-game fixture. Native browser task scheduling needs inspection if (program) remains dominant. Do not change _ask, equivalence, cached timestamps, presentation, state, or authority guards based on this profile, and do not resume this CLOSED runtime.
Production adapters/tests were unchanged; matching 51 Quake and 42 Hex controls were reused. No production old-fail/new-pass or full-build qualification is claimed. Cleanup verified all 26 supervised PIDs nonlive, actual joins and EOF, finite 600/630-second independent supervision, removed owned profile/display/socket/auth and empty task child, and preserved proof files and the assigned Node cache. Runtime minimum sampled home headroom was 1,592,803,328 bytes; continuous headroom is not proved. Box ID, expiry and stop receipt are null, state NOT_CREATED.