Hexen II software browser encounter, 2026-11-18

Actual software rendering, living movement, incoming damage, two visible spiders, trusted Mouse1 release and ordinary Quit with exit code 0 are proved. An attack while alive, enemy hit and enemy defeat remain unproved. The last reviewed target frame had health 5; Mouse1 DOWN arrived 11.906 seconds later. The next whole frame showed health 0, a rolled death view and a loading overlay. That interval cannot establish an alive attack or attribute damage to an enemy. No further combat input was sent.

Evidence: scratch/runs/20261118-hexen2-software-living-enemy-consequence/. This complements the WebGL encounter recorded in hexen2-dialog-release-browser-gameplay-20261118.md; its enemy-hit/kill and document-mouseup limitations remain unchanged.

Immutable runtime and actual renderer

The memory-only Git server used published commit 0b55b3c2f1be569f8799ee1390ebfd8c620cb39b. Compiled, served and recovered executing guest/render Worker module SHA-256 is a6002d5dbcfeb534a0eadd223da616e525aa1542dad2bde381fe54e00ee862f8 (2,418,710 bytes). Foreign worktree overlays were excluded. All 18 manifest, media and registered-control references remain unchanged. The registered glh2.exe route and startup Flag were retained; the unregistered h2.exe route was not substituted.

The browser backend selector is ?gl-gps=software, not merely the historical gl-renderer spelling. The actual owning gl-gps endpoint requested software; natural completed presentation receipts contained positive nativeDraws, nativeBytes and presents, and null glRenderer. These are the production software executor's WAT rasterizer counters, forwarded by the pinned render Worker adapter. Both actual executing Worker modules were independently read through CDP. Whole headful X11 PNGs show the resulting authored triangle and original textured menus/courtyard/chamber. A selector label alone is not the renderer evidence. Normal guest/render Workers, clocks, focus and polling were preserved; Chrome viewport emulation was disabled.

Authored controls and input boundary

The authored PE opens a real blocking MessageBox before creating its GL window. Balanced Enter independently released against the exact process/document owner before separate GL readiness observation. The initial scope then closed on an incorrect harness assertion requiring a nonexistent software backend stat. Its module and native-rendering receipts were retained; it received no further endpoint requests. A later capture command was refused locally before dispatch. The assertion was corrected offline, with no runtime modification, and a distinct fresh authored scope passed the full control.

The fresh control retained the visible notice, actual software triangle, Mouse1 changing the background from blue to red, neutral physical state after UP, ordinary Escape exit 0, and actual process/document replacement negatives. The raw startup dialog rectangle has RGBA SHA-256 00dd825eb41e2087b9ac81374142ac7f021b34a6356567e8e5f67ff815de9e19, matching the visible notice in both authored scopes.

The run-local sender prearms an independent child bound to the exact process and document, then sends ordinary DOWN and UP through that child's one CDP session. This is a qualified input route, not a causal fix for the prior run. The current registered CTRL binding is crouch, overriding default CTRL attack; Mouse1 is the unchanged configured attack route. No keyboard attack binding was invented or changed.

A document-only event ledger misses mouseup because production windowMouseUp receives it in window capture and calls stopPropagation(). A passive window capture observer, installed before the next gesture, retained trusted DOWN and UP under pointer lock in the authored control. The original also retained its trusted window DOWN/UP pair and passive Mouse1 state 0. The original's document ledger still lacks UP. These fresh observations do not recover a missing event from the closed predecessor or establish guest attack consumption from transport acknowledgments. No production helper/runtime fix, native/browser old-fail/new-pass repair, or code-build claim is made here. JavaScript syntax and the existing native process-owner focused tests passed; actual browser replacement children refused stale owners before arming.

Original observations and ordinary Quit

Paired native menu inputs selected Single Player, New Game, Paladin and Apprentice. Model meshing completed naturally. living-courtyard.png shows upright fists and health 85. A right turn and W holds visibly moved up the ramp and through the skull doors. The player reached the chamber wall and received incoming damage while turning: health 74, 51 and 34. No attack was sent during those views.

Unchanged default Z look-down exposed two close spiders in floor-target.png, with health 5. The spiders were above the crosshair in that steep downward view. The subsequent Mouse1 attempt has a trusted browser pair but no retained alive attack animation, enemy-specific injury or corpse. combat-boundary.json records the exact screenshot-to-DOWN interval and proof limits. The next view was death plus loading; later the native menu appeared over a reloaded courtyard at health 81. The attempt may have been consumed as a death restart. There was no intentional replay or second combat attempt, and no runtime cause for a failed living attack is isolated.

First Escape released pointer lock; the second eventually opened the native main menu after loading. ArrowUp and Enter reached the credits confirmation with “Press Y to Exit”. Paired ordinary Y (virtual/native keycode 89) released to physical state 0. ordinary-quit-exit.json records running false, processExited true and exit code 0; the whole terminal PNG shows shell Ready. The run closed immediately afterward without further endpoint requests.

Bounds, cleanup and recovery

Predeclared windows were startup 180 seconds, each authored scope 300 seconds, original 1,200 seconds including a 240-second Quit reserve, and cleanup 20 seconds. Independent watchdogs bounded runtime plus cleanup. Holds were at most 1,800 ms, release fallback 5 seconds, CDP 15 seconds, whole capture 10 seconds and review 90 seconds. No window was extended. The original ran 09:38:14–09:47:17 UTC. Its sampled minimum free space was 1,508,761,600 bytes, above the 1.5-GB floor.

Browser, display, release children, supervisor, launcher, watchdog and enclosing invoker waits are retained with actual child close and both EOFs. Containment reported no leaks. Git remains the source archive; relevant binary diffs, module/media manifests, run-specific helpers and exact recovery instructions are retained. PNG compression is lossless, with decoded-pixel equality and exact original-PNG regeneration checked before replacement; screenshot-time hashes remain recoverable. There are no source snapshots, media copies, agents, workers, boxes, Python, Wine, deployment or held-title actions. Foreign UT, AoW2, ReVolt and held BloodII work was preserved. Box ID null, state NOT_CREATED, expiry null, stop receipt null.