Uplink: native PageDown lowers aim; combat still unproved

Run: scratch/runs/20261112-halflife-native-look-pause-headcrab-1905/, with -control/ for the private lifecycle and blank control.

The actual emitted finite input consumer now supports End VK35, PageDown VK34, PageUp VK33 and Pause VK19. Ordinary native PageDown visibly lowers the restored elevated camera. End centering and Pause/resume are not qualified. The headcrab kills the player before the newly exposed enemy can be centered and attacked. No outgoing hit, kill, safe room traversal, new save or successor export is established.

Controls frozen before allocation

Only the run-local input.cjs key table changes. The final input SHA256 is c775c93ccbff8b74efb70fa23d99b6a621f498c7a773363b3c82f661f24fc303. All four special keys omit injected text. The unchanged actual execute consumer is extracted from the emitted browser.cjs and exercised through the numbered request/ACK transport. Twenty-one cases pass both locally and through real Chrome trusted native events echoed by a real Worker: four successes, per-key UNKNOWN and failed-UP cases, partial chord failure and eight rejected keys with zero admission/effects. Every admitted case retains its armed-before-down receipt and independently attempted releases. The first test attempt used an uppercase capture label and was correctly refused before effects; its evidence remains under attempt1-test-End/.

Seven unchanged readiness, strict22 binding, capture, delivery and transport helpers and four retained proof reports are independently checked by SHA; their expensive unchanged matrix is reused. final-freeze.json pins767 files; the manifest SHA is d399308336ba4b62f0d1ff67f16307f50c8413bcf6458f1b598a2618921a570c. Local and remote final pin/import gates pass before original start. No runtime helper is edited after allocation. Git commit/source diff and helper patches are retained instead of a source archive.

The original loose default.cfg SHA b625391874c7beca74b9d2f643dfbc5343062bd52c7d5c8780289fee2072666b supplies the static binding facts; those facts are not substituted for live camera or pause evidence. No configuration write supplies these tests.

Whole13 Load and native camera evidence

The complete original checkpoint from the earned rear-panel passage is independently reverified:758472 compressed bytes, SHA256 df8d61b2fcdd285a6a28e2e180930e6ee60eff388cd9bf31c557d3409dd18e51, 13 strict-base64/size/SHA-checked files totaling2342063 decoded bytes. The005 save remains336143 bytes, SHA256 5d02d6bc10c55a7502acf69a83f7bde3b6aab8fe9c07e728d562dc83ab6b7bdb. All13 members, including both maps' auxiliaries, pass fresh-browser VFS verification together before native Load; there is no mixed checkpoint.

Native menus select the bottom03:18 row in n12-select318, then Load in n13-native-load. After the actual loading screens, n17-loadwait shows the elevated corridor/crowbar at 35/0, already taking damage. n18-bound binds the living actual CPU and owning SoftwareGL module to 04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17; the full native frame now shows 30/0. Save-time40 is not treated as current health.

n19-pause sends ordinary Pause for120ms. n20-pause-observe retains30/0, but n21-center subsequently shows15/0 and fresh damage. This does not establish useful pause/resume. End400ms in n21-center leaves the view elevated; acknowledged delivery is not centering proof. The possible effect of original mlook is not diagnosed by this run.

n22-lookdown, after ordinary PageDown1800ms, visibly replaces much of the ceiling with the lower wall and floor. Health is10/0. The ordinary backward retreat in n23-retreat changes position and exposes a nearby floor headcrab left of the center, at5/0. These full native frames prove downward view and fresh movement, without assigning an exact pitch, guest hold duration or hitbox. A short450ms left turn in n24-align-crab ends in the 0-health rolled death view with the headcrab, preserved in full. No attack is sent against an uncentered or unseen target.

Disk interruption and exact terminal cause

The next ordinary Escape/menu command is twice refused locally before remote effects because actual /home free space is below1.5GB. The minimum owned sample is1470066688 bytes. Free space fluctuates above and below the floor while other agents share the host; no continuous compliance is claimed. The worker compresses selected own JSON with exact raw/gzip SHA and roundtrip records, without changing reserves, touching another worker's files or gating on RAM.

During that cleanup, review JSON is compressed and then restored for the unchanged caller. Restoration changes its file times. The caller selects the latest local capture by mtime, so the eventual n25-death-menu request incorrectly names n21-center, while the remote consumer's latest is n24-align-crab. The actual retained first failure is assert(latest&&c.reviewRef===latest.label), before admission/input. This is an exact-reference mismatch, not a proved age-limit failure. Earlier commentary called it stale; retained artifacts correct that description. No n25 input admission/pulse or ACK exists, and no replay follows.

The original guard terminates its owning process tree. Normal post-death Load is not completed in this run. There is no closing save export because the session ends by terminal cancellation. The authoritative parent whole13 file is rehashed unchanged and remains the recovery source. The prior run's normal post-death Load is separate evidence and is not relabeled as a result here.

Before another allocation, replace mtime-based local selection with the exact latest acknowledged capture identity and test restoration/reordering locally. Then use the proved PageDown delivery earlier after Load and promptly retreat or advance before turning left. A longer native Pause test would still need actual pause/resume evidence before relying on it. Do not repeat blind melee, the old crate or the crouch loop.

Verification and lifecycle

terminal-verification.json passes287 recovered original references, 25 full PNGs with CRC/SHA and all22 within-capture identity fields, seven captures against the living gameplay anchor,211 HTTP pins, 17 balanced pulses (six key pairs/11 click pairs),45 trusted native events and24 numbered ACKs. Release attempts are -1..3ms from planned times. End, Pause and PageDown each have an actual native DOWN/UP pair. PageUp has the local Chrome/Worker proof only, with no live-game effect claim.

The independent two-hour watchdog is armed and observed alive before the single private creation. The fresh blank Chrome control passes23 recovered SHA references and132/132 custody tasks, actual outer wait0 and both EOFs. Original cancellation records owned-process exit-15, actual outer wait1 and both EOFs; 363/363 custody tasks reap with no remaining/unreaped tasks. This is complete cancellation cleanup, not an ordinary clean game exit.

All unique original references are recovered and hash-verified before normal delegated box stop. bx_6rbx9a95 is archived, expiry null. Initial expiry 21:07:13.230Z. Normal stop stop_38b853f40477ceb5 completes 19:16:21.252Z, forcedfalse/errornull. Actual local manager and watchdog both join0 with both EOFs. No box is deleted, resumed or left billing idle.

No production source change/build, original EXE/PAK/media modification, console, guest position/health/AI/clock/memory writes, forced guest calls, deployment, other worker, full27, parity, FPS, audio or broad crash-free claim results. The remaining blockers are living aimed combat/safe onward progress, unqualified End/Pause effects and the run-local latest-capture selection flaw.