TR2 capture delivery and review admission — 2026-11-10

tools/tr2-review-admission.py replaces four serial signed-file downloads with one immutable capture packet. The command response returns its SHA/size; one download contains the original PNG, capture, state and binding bytes. The receiver verifies the packet and every member, checks PNG/capture binding, and fully decodes PNG with CRC before display. The calling tool displays that image in the same invocation that observes command completion. Individual model review remains a subsequent explicit command argument. No review is generated by the helper, and original at, completedAt and owner identity are preserved. The helper never refreshes or retries after failure.

This is title-only transport work. The generated browser caller is byte-identical to tools/tr2-visible-target-margin.cjs output from f5b4c9532. Its 12,000 ms freshness fence, 300 ms requested navigation limit, 400 ms release-completion limit, visible-enemy-before-shot requirement, unchanged-canvas fence and sticky owning cancellation remain unchanged. Shared runtime, render and startup providers were read-only. The accepted native fixture was verified by reference and not repeated.

Evidence and controls

Evidence: scratch/runs/20261110-tr2-review-admission-latency-0432/. Git SHA, relevant binary diff and owned new-file patches identify the tested source. Original media is selected by existing path and SHA; no source archive or media copy is retained in the run.

The preceding 0415 run captured world-first at 04:22:56.690Z and completed it at 04:22:57.686Z. Its command returned at 04:22:57.809Z; the next command started 16,837 ms later, after four serial file downloads and model/tool work. Admission refused age 18,116 ms. That evidence does not isolate the old downloads' individual durations. The new actual delivery removes six serial network operations (three signed-URL requests and three GETs); it does not remove model decision time or owner identity validation.

Before allocation:

One original and the remaining limit

One UUID-leased original used 654 pinned inputs and the unchanged accepted fixture. All uploads joined before assembly/SHA verification; preflight passed without starting Chrome. Startup Graphics was personally reviewed: Hardware 3D, Full Screen, 640×480 High Color. Ordinary Enter100 accepted those settings. Its dependent command used the original startup capture, without refresh.

World-first was captured at 04:35:36.136Z, completed at 04:35:37.138Z, SHA256 5c375fc618b3f662c65e447a399810069c1a230771ccbd180223c0a6eade3e8c. The packet download/full validation finished at 04:35:37.512Z, 276 ms after the remote command returned. The PNG was individually reviewed: upright Lara in the rising alley, brick left and green wall right, no visible enemy.

The distinct first ArrowLeft300 was admitted at 04:35:45.368Z, original image age 9,232 ms, within the unchanged 12-second fence. Release completed 302 ms after start. The resulting left-first image, captured 04:35:49.274Z, showed a shifted alley projection; its game region changed 223,849 pixels (56.3%). This establishes changed visible output, not guest keyboard polling, a frame rate, physical GPU parity or encounter progress. No visible enemy was identified. Packet verification completed 04:35:50.468Z, 233 ms after that command returned.

The next individually reviewed ArrowLeft300 request hit the first failure at 04:36:04.446Z: original image age 15,172 ms. It was refused before dispatch; there is no second-left admission, hold or subsequent screenshot. The repair removed avoidable transfers and admitted one previously blocked step, but the subsequent model/tool decision interval still exceeded the fence. There was no Space, Control, observer arm, ammo change, target identity or target-health observation. The visible-enemy/native-shot objective remains open. No new box or repeated original is authorized by this finding.

First failure latched owning cancellation at 04:36:04.449Z. There was zero later guest protocol, input, capture, refresh, retry, quit, extension or second native run. The outer parent actually waited status 1 with separate stdout and stderr EOF. Kernel custody records 173 owned / 173 actual reaps, complete true, zero remaining/unreaped/unresolved entries. Normal Browser.close and a Chrome exit-zero are not claimed; CLONE_UNTRACED remains unaudited.

Retrieval and exact stop

All 85 unique artifacts / 2,336,843 bytes were independently reread and SHA/size verified, including full decompression/CRC of 14 gzip streams and full decode/CRC of three PNGs, before the sole delegated unforced stop. Selected inputs stayed unchanged. Runtime modules matching selected bytes use verified references. Staged chunks and Chrome profile cache are excluded.

The independent watchdog was observed before allocation, UUID 2734ae16-925c-43fb-8159-fd4313c1bf89; immutable deadline 04:59:41.944841Z, initial API expiry 04:59:41.325Z. The supervisor actually waited it with return code 0 and separate stdout/stderr EOF. No clock, tier, deadline or freshness setting changed.

Box bx_xvaerjjt is ARCHIVED, expiry NULL. The sole unforced stop receipt stop_e2da66f82d297b62 completed 04:36:36.464Z, error=null; stoppedAt 04:36:36.447Z. Other boxes were untouched. No local Chrome, nested workers, guest writes, artificial native calls or public deployment.

Further work must shorten actual model/tool decision latency within the existing fence; this run supports no further native retry. Worker releases STOP after publication.