HL2 agreement capture delay, 2026-11-17

Retained genuine installer evidence: scratch/runs/20261117-halflife2-published-async-trace-normal-install/result.json. Welcome and Internet Next succeeded; agreement capture exceeded unchanged 3000 ms and permanently closed UNKNOWN. Its late whole agreement PNG proves rendering, not acceptance or installation. No old UNKNOWN endpoint was reused.

The three retained desktop trace streams contain 388 nonempty slots; every 4096-byte slot checksum verified. Main trace poll18045 completed before a 3301.885454 ms interval ending at snapshot18046 submission. Sixteen /proc samples within that interval account for 1481.123451 ms main CPU and 1620.307614 ms scheduler wait; broker threads slept. This is not evidence of an outstanding IPC failure. No historical IPC attribution is made.

test/test-halflife2-capture-workload.cjs creates an ordinary owner, native Internet modal, genuine native Next dispatch and subsequent native dialog containing exactly 453207 authored bytes in a readonly multiline Edit with scrollbar, sibling controls and comparable geometry. It uses the real CLI compositor, poll and PNG capture path. Steady already-laid-out capture succeeded (98 ms internal); baseline transition reproduced permanent capture3000 UNKNOWN and a late whole agreement PNG. One malformed callback fixture and one premature listener preflight were retained as failed authored setup attempts.

Exact retained Edit module 53c2791da77f4e751c92b77a02f1503ea23741b695db1073230f1b4116eaba2e was reproduced byte for byte from pinned sources; a separately compiled name section maps the profiler indices. A bounded authored transition profile attributes 2721.68 ms to gdi_bitmap_font_selected under host_measure_text / edit_layout_build_page. Profiling overhead shifted this run's timeout to the guarded input's final snapshot, so it is localization evidence, not an identical capture timing measurement.

The Edit builder measured the entire growing prefix for every next byte, repeating font and glyph work. For the undecoded bitmap path, width is exactly the sum of independent character advances. The repair accumulates those advances once per byte within each line. Decoded codepages retain the original counted-prefix path. No persistent cache, control predicates, native ACK500, UP1750, capture3000, one binding, first-error/no-retry or UNKNOWN transport rules change.

The sealed adapter recompiles only this WAT delta against the exact retained graph. Fixed module SHA256: 928f8bd42c76b6435ec59d8b289b5a40a28da8b597a6e98e378b467311f11abb. Native fixed transition passed the aggregate guard (582.535 ms), and the agreement PNG capture completed internally in 85 ms, with the same 8994 native visual lines. Personally inspected whole agreement PNG includes readable text, scrollbar and labeled Next, Cancel and Back. A later authored exit attempt expired its quit review locally and latched UNKNOWN without sending input; it does not qualify the complete workflow.

Direct native-WAT regression validates all453207 bytes, exact first10 line offsets/lengths and navigation past the384-entry scratch page for both baseline and fixed module. The direct harness stock font produces10792 lines in both; timings were3568.73 ms baseline and147.29 ms fixed. This is an Edit algorithm change, not an interpreter/JIT variant. Canonical locked build and focused validation are retained under scratch/runs/20261117-halflife2-agreement-capture-delay-cause/.

Repair commit ccb616c3c was immediately strictly pulled/pushed. Fresh published authored diagnostic qualification passed, then the complete one-bind workflow passed Welcome Next (346.778 ms) and reached SubscriberAgreement on Internet Next. The latter independent release HTTP call timed out (physicalUpACK false), so unchanged UP1750 failed, UNKNOWN latched and the owner stopped. No later HTTP was sent, including the subsequent locally refused capture. This is a distinct unresolved release failure, not a capture-repair success claim for the complete workflow. Original-run prerequisite therefore failed: original invocation count remains zero, and the allowance is unused. No acceptance or installation is proved.

Next work is to diagnose that retained published authored UP failure and qualify the complete released workflow before any separately authorized fresh original. Free disk also crossed the1.5GB floor due concurrent writes after this run had already stopped; no further native execution was started. No boxes created. Source recovery uses Git plus binary patches; existing media is referenced by path and SHA256 rather than copied. All fixture runtimes have actual native/owner/broker/watchdog/outer waits and EOF records; host cleanup is distinct from native Quit.