Forsaken capture ACK and bounded diagnostic admission, 2026-11-12

The run-local capture/ACK orchestration is repaired and passes combined controls. The original-game attempt stopped before provisioning: a local helper changed between pinning and allocation, so whole-input validation rejected its hash. No original was launched. The combat-white cause remains unproved; there is no production rendering change or changed-original before/after claim.

Evidence is scratch/runs/20261112-forsaken-nonblocking-diagnostic-white-1550/. The first-white handoff and owner/enemy evidence remain the source of the original observations. Menu execute-buffer opaque-state restoration does not establish the combat transition. Postwhite additive submissions do not establish the upstream writer.

Repaired path and controls

browser.cjs completes and publishes the full PNG, binding and capture metadata without awaiting diagnostic telemetry. Its shared cycle.cjs command path lets the existing transport publish ACK, then drains diagnostics serially. A separate admission gate stays busy until that drain has a known outcome. The requester waits before publishing the next command or calculating its command budget. Periodic drains hold the same writer lock used by request publication. Thus a periodic drain cannot consume an already-published command's work budget.

Diagnostics have a separate 60-second total lifetime; each existing CDP request retains its 20-second response bound. An unknown response is terminal, including after a successful capture ACK. No later input, capture, quit or retry is allowed. A settled diagnostic exception records an explicit drop and disables subsequent diagnostics. Request-admission failure records sticky closure before publication. There is one CDP owner, no diagnostic request queue and no abandoned request used as permission for the next physical command. Native balanced release, complete capture, identity, hash, freshness and accepted command deadlines are unchanged.

CPU retention is 32,768 rows per drain, with explicit overflow counts. Each drain is compressed immediately, with a 24 MiB raw payload check and a 32 MiB total compressed retention cap. Reaching the total cap produces an explicit drop and disables telemetry. These bounds do not certify uninterrupted original evidence.

combined-cycle-test.log exercises the exact emitted browser execute/capture/ drain functions and the same command-cycle module called by the browser, using an offline CDP/identity peer and a complete 1024 × 712 PNG. The trace payload is produced by the real built module's x86 Device2/3 public-COM pixel fixture: 1,701 entry/return/changed-state rows, zero drops. A deliberate 21-second drain outlives command 0's unchanged 20-second work deadline after successful capture ACK. Command 1 remains unpublished until the drain completes, then gets its own full budget. Exactly one Enter down/up pair is observed. Known diagnostic drop, unknown response and terminal no-replay controls pass. This is combined orchestration validation with an offline peer, not real-browser original proof.

periodic-test.log separately proves predrain serialization and budget timing. retention-final-test.log proves the total cap emits a known drop without more CDP, and preserves an already-terminal owner. The initial retention control caught a disabled-observer early return bypassing the terminal check; the check now precedes that return. CPU overflow controls, inline owner/rejection controls, three native-release tests, thirteen API custody tests and locked tools/build.sh passed. The immutable module remains SHA-256 04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17.

First blocker and lifecycle

I added lifecycle assertions after pinning lifecycle.py. Provisioning correctly rejected that stale pin before its first upload or remote command. The terminal record is preserved; there was no replay or second allocation. The controller now reruns the actual byte/hash/import preflight immediately before allocation, rather than trusting an earlier pass log. The stale-input check fails and the repinned final preflight passes. This preparation fix has no original retest.

Source is recorded by Git SHA plus relevant tracked/untracked patches. Final helpers are preserved in owned-inputs.patch and individually pinned in runtime-pins.json; no source tree or media archive was made. Existing original media paths/hashes were verified. Local /home remained above the 1.5 GB floor.

Private box bx_bzn68cwr had an independently observed two-hour watchdog armed before allocation, initial expiry 2026-11-12T17:55:44.945Z. There were zero remote outputs or original commands to recover. Delegated normal stop completed at 2026-11-12T15:56:48.649Z, receipt stop_c4d630ac6c630370, forced=false, error=null. Final state is archived, expiry null. Lifecycle client, adopted owner and independent watchdog have actual successful waits and EOF evidence. No Chrome/runtime parent existed. Do not resume or delete this box.

Next: use the finally pinned, tested combined path for one fresh original current module/owning-WebGL run, retaining the CPU public-RS/changed-state and natural GPU pixel transition before first white. Compare damage, respawn and opaque/additive transitions, preserving every overflow/drop. Patch the emulator only if that trace proves an upstream cause. No guest writes, forced calls, FPS/parity, full-27 or broad crash-freedom claim was made.