Forsaken raw-state observer startup blocker, 2026-11-12

The requested original first-white boundary remains unresolved. The new observation-only helper passed focused local regressions, but its first worker injection in the actual original browser run timed out before readiness or any game capture. Runtime.evaluate had an unknown outcome; it was not replayed. No original execute-store/enqueue/replay/describe witness was obtained, no runtime defect was proved, and no runtime repair was made. Stop at this concrete startup-observer blocker; do not infer a blend, clear or texture fix.

Read the previous original first-white result at a04a489fc for the retained GPU21645 white transition and subsequent additive GPU21660 despite CPU opaque restores. That evidence and its limits are unchanged. Prior root review remains pending; this task adds no acceptance or broad compatibility claim.

Scope and tested observation

Evidence is in these two local runs:

Only run-local helpers and this note changed. Git remains the source archive: each run records its exact main SHA, relevant tracked/untracked binary patches and owned-helper patches as lossless gzip, with raw hashes in patch-recovery.json. The second source base is 42b48e0814b0e7221c6bb75866e9ffae81bbccfd. Existing original media paths and SHA-256 values were checked without source-tree or media archives.

The helper observes changed-state pre-store hooks and raw memory at the next natural import boundary, live state at encoder entry, owned packet bytes after copy and before publication, source packets at replay, replay workspace, and raw state before/after description. A next-boundary read is an interval witness, not an immediate hook after the store. It records device/state addresses, raw RS1/7/14/15/19/20/21/27, a full render-state-region FNV checksum, buffer/offset, sequence and ordinal. The checksum is auxiliary, not a cryptographic proof or a substitute for retained raw fields. Original methods retain arguments, return values and error identity. CPU post-store and replay address reads use the existing pure JS translator. The enqueue observer additionally queries the encoder's existing guestToWasm translator, which can call its pure translation export. It invokes no guest setter or application API.

Focused tests passed before the second allocation:

Each attempt froze all remote helper/data pins and passed the actual local Transport.provision whole-input seam, a stale-final-pin rejection, syntax and transitive import preflight before allocation. Lifecycle rechecked current bytes immediately before creating the box. Provisioning verified all 1,560 remote paths including final pin/deadline metadata; remote import preflight passed before the second Chrome launch. No RAM gate was used; sampled actual /home free space remained above 1.5 GB.

The original CPU/GPU caps, 24 MiB drain payload bound and 32 MiB compressed retention bound remain; the new observer has a separate 32,768-row cap and explicit drop/error counts. No actual original diagnostic drain completed, so there is no original zero-drop or state-preservation result from this task.

Actual blocker and shutdown

The first attempt was stopped before Chrome after noticing that the observer checked the guest pointer against the translated direct-window bound. The regression had not covered the previous original's high device pointer 0x08222048. Its one unique remote output was independently downloaded twice and hash-verified before normal stop. The final helper instead uses the existing JS direct/DIB/sparse translator and tests the high-pointer case.

The second browser navigated to the original threaded WebGL route at 2026-11-12T16:34:50.797Z. Its HTTP build checks passed and the recovered page module matches the normal hash above. That does not establish guest/renderer instantiation or an actual WebGL owner. Ordinary blank-page native preflight passed before navigation; no game input command was published.

In recovered/webgl/protocol-timing.jsonl, worker session 571C6A8AA104ED87EC20D66441D1C832 completed Debugger.enable and Runtime.enable at 16:34:51.698–.699Z. The new boundary script is the next evaluation in the frozen attach handler, before the old observer and Runtime.runIfWaitingForDebugger. At 16:35:11.705Z, worker-errors.jsonl records CDP UNKNOWN Runtime.evaluate; no replay; first-browser-failure.json records the terminal failure at 16:35:11.903Z. There is no successful resume/readiness, startup PNG or raw diagnostic drain. The exact internal reason for the evaluation timeout is unproved. Loading a library or touching worker startup globals before resume is a hypothesis, not a demonstrated runtime defect; the retained target URL is insufficient to establish a renderer-specific cause.

The browser's existing terminal-failure path closed Chrome (its failure branch uses SIGTERM). No game normal-quit or successful Browser.close command was observed. Chrome's actual parent wait is code 0 with both EOFs. The browser runner's actual parent wait is status 1 for the observer failure, with both EOFs. A later owning-PID semantic-stop command failed before sending a signal because /proc/7910/cmdline no longer existed; semantic-stop-outcome-readonly.json retains the known exit-1 cause. That command made API custody terminal, so no later runtime commands were sent.

All 26 unique remote outputs were downloaded twice independently and rehashed, including the module and actual parent/EOF records. An outer API-command status GET attempted after archival returned HTTP409; its final status/EOF is therefore unobserved and is not inferred from the nested parent wait. No box was resumed to fill that gap. Local lifecycle client, adopted owner and independent watchdog all have actual zero waits and EOF evidence.

Both independent two-hour watchdogs were observed alive before allocation. Delegated normal stop followed unique recovery, with these final states:

Box Initial expiry UTC Final state / expiry Normal stop
bx_kmv8ha2n 18:26:47.802 archived / null stop_4755503a04689102, completed 16:30:15.420Z
bx_3czfgkvx 18:31:55.842 archived / null stop_bab4a866501599b2, completed 16:36:32.224Z

Both receipts have forced=false and error=null. Do not resume or delete either box. Large closed local logs/traces are losslessly compressed with raw hash recovery records; runtime input pins remain intact.

Next work must first prove the actual paused-worker injection/resume seam using the final observer, without an unknown command or replay, before allocating for another original menu/combat trace. Only then can these raw state witnesses locate or exclude a first divergent boundary. No blend/clear/texture workaround, guest change, workers, public deployment, source/media archive, full-27, parity, FPS, audio, enemy-defeat or level-completion claim was made.