Heretic II crate load: observed GL error and context boundary — 2026-11-12

Fresh current04 SoftwareGL loaded the exact whole-four crate0400, showed an upright saved crate-recess world at HUD health 31, completed fresh backward movement, and remained upright through a separate released interval. Fresh WebGL again called native ExitProcess(8) during normal loading. The new permitted GL observations do not identify a runtime violation explaining that decision. No production patch or new save was made.

Run: scratch/runs/20261112-heretic2-runtime-boundary-1910/. Authoritative recovery remains parent 20261112-heretic2-current-paired-crate-onward-1110/RECOVERY.md and 20261112-heretic2-current-load-exit-cause-1135/RECOVERY.md. All four preferred originals remain together under 20261112-heretic2-north-opening-health-crate-0400/save-inputs/crate0400/: server.ssv 2,752 B, game.ssv 62,728 B, demodocks.sav 1,395,332 B, demodocks.sv2 90,112 B. crate-manifest.json retains their exact hashes. Neither the separate living 0237/0136 fallbacks nor failed auxiliary sets were substituted or changed.

Runtime source stayed frozen at 8a834a3dc7affc8c6e53245beab56e87a32b81e8; actual guest and owning renderer modules in both arms were 04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17. Checkout entry was c458e7c5105f13a6a343567114f218eec6220a86. Original registered media and unchanged qualified helpers were reused by SHA. The final 2,423-file closure, changed observers, shared-memory forwarding tests, return/exception identity tests, observer-fault isolation, 22-field null/change refusals and normal-load-only console schema all passed before box allocation. Actual provisioning joined 0 with both EOFs and verified 2,425 pins, including manifest/deadline, before either original browser started. The final frozen-file check also passed. Later helper files perform only terminal inspection, recovery, lifecycle accounting and evidence checks; the runtime observer was not edited after allocation.

The new renderer observer forwards original calls once. It records existing WebGL getError() results without issuing an additional error query, bridge glGetError results, context creation/current transitions, frontend error transitions, exceptions and context-loss events. Presentations are sampled every 60 calls; this is not an exhaustive presentation-return trace. Its native-import wrapper produced no direct GL-query return records. That absence is consistent with the actual runtime contract: the WAT encoder submits bounded streams through gpu_gl_call(0x10001, offset, length); ordinary direct GL opcodes are rejected by lib/guest-rpc.js. The renderer replay records supply the actual observed GL-query results. The synthetic forwarding tests alone are not evidence of guest-visible query writeback.

Both arms created context 1 and made it current successfully, with no later observed context switch/deletion before the exit or ordinary closure. Each arm made one bridge glGetError query, returning 0 at startup. WebGL's six existing native WebGL getError() reads also all returned 0 at startup. Every observed error query predates the save mount; neither arm queried the backend error flag during normal checkpoint loading. There were no recorded frontend error transitions, context-loss/restoration events, renderer-call exceptions or observer errors. This does not establish that an unqueried backend error flag was clear during loading, nor explain the native exit.

Actual ownership comes from GL endpoint 1, live context objects, renderer target and epoch, recovered modules, completed presentation and full screenshots. SoftwareGL has software:true/backendSoftware:true; its separate legacy WebGL endpoint is irrelevant to that identity. WebGL has both flags false, WebGL 1.0, and the actual reported Mesa/Intel renderer string. No physical GPU claim follows from that string. All 30 full captures retain stable, non-null 22-field bindings before/after capture and before admitted input. The internal phase label remains startup; actual GL ownership and all identity fields were verified directly, without treating that label as gameplay qualification.

Each browser followed native Main → Game → Easy, reached original upright docks with HUD, mounted the exact four-file manifest, verified whole-file VFS hashes, and visibly reviewed/submitted only load crate0400. SoftwareGL's completed saved-world, fresh-backward and released-survival PNGs establish its living control. WebGL's fresh-start rat reduced HUD 100 → 98 → 92 before console entry; those captures are preserved, with no health write or claim of identical pre-load gameplay timing. The loaded save bytes, media and runtime closure were identical across arms.

WebGL's API-tail timestamp places native exit 8 at 19:33:03.672Z; the host console received it at 19:33:03.727Z. The proposed console-close command started its admission reads while the guest was still loading, then received Session with given id not found at 19:33:03.728Z. There is no admission or pulse receipt for that command, and the last dispatched input remains the normal-load Enter release. Its retained operation exited 1; nothing was replayed. Chrome closed 0, and the runtime parent actually joined 1 with both EOFs. SoftwareGL's runtime parent joined 0 with both EOFs. The familiar final API-tail successes, save coverage and optional model misses are not new causes.

The first remaining permitted witness is the final completed owning-GL presentation retained before teardown, bound to the native exit notification. The final API tail records wglSwapBuffers returning 1 at 19:33:03.667Z, five milliseconds before the exit call. The last full PNG is the earlier black loading/console frame. A late CDP screenshot cannot recover a retired renderer. A future prequalified observer should retain the latest completed compositor/ GL presentation pixels and recover that frame on native exit, without changing guest calls or interpreting private guest structures. It might preserve visible native diagnostic output; this run does not claim that such output exists or that the final frame explains the exit. Do not repeat startup error reads or the closed save/model/API-tail findings as a cause or a fix.

check-evidence.py verifies 364 remote references / 276 unique raw hashes, encoded/raw hashes and sizes, 30 PNG bindings, exact save mounts, stable full 22-field identities, GL observations, balanced trusted keyboard edges (SoftwareGL 56, WebGL 38), runtime-parent waits, profile cleanup and normal stop. compressed-evidence-manifest.json and retrieval-verified.json recover each original. Git SHA, relevant binary diffs and owned-helper patches recover source; no source/media archive was made. No guest reverse engineering, disassembly, extraction, patching, substitute DLL, forced guest call, config/ clock/state write, onward qualification, new checkpoint, FPS/audio/parity or full-corpus claim is made.

Own private box bx_uh4ghfzm had an independently observed two-hour watchdog before creation and initial expiry 21:13:51.567Z. Unique evidence was recovered and hash-verified before normal stop at 19:35:42.087Z. Final state is archived, expiry null; receipt stop_efe142b2325cc57b completed unforced, error null. The fresh stop subprocess actually joined 0 with both EOFs. The original local manager exec ended 143 without its parent wait file; the watchdog survived reparented, observed the completed stop, and exited. Original manager/watchdog parent waits remain UNKNOWN. Both browser parents' actual waits are retained separately. Shared disk-floor dips held launch during provisioning; actions resumed above 1.5 GB. No reserve change, RAM gate, additional worker or public deployment occurred.