Aranna independent release repair; allocation held at disk guard

Ordinary snowy-world movement or attack remains unproved. The run-specific mouse adapter was repaired and exercised in actual Chromium. Complete final retained-runtime qualification remains incomplete: its local control reached guarded acquisition and observer arming, then exited normally at its fixed shutdown time before the target action. A fresh local control was refused before Chromium launch by the unchanged free disk >2 GB guard. The later measurement was 1,926,557,696 bytes at 11:28:23Z. No original box was allocated.

Evidence is scratch/runs/20261107-aranna-independent-release-gameplay/ (R). R/qualification-refusal.json records the prelaunch refusal and scope. This is a harness/environment result, not a demonstrated emulator mismatch. The old source-flow release blocker was repaired rather than merely reported again.

The resolved adapter in R/input-adapter.cjs replaces exactly one retained serialized button branch with R/balanced-release.cjs. It schedules mouse-up before invoking or awaiting mouse-down, at min(start+60 ms, release deadline). Down and up acknowledgements have independent bounded waits. Up is attempted once even if down rejects, remains pending, or receipt persistence fails after arming. Any rejected/missing ACK or persistence error produces UNKNOWN; the existing adapter closes the queue without recovery or replay. A successful receipt requires both exact transport promises to acknowledge. This does not claim that a transport failure guarantees physical button balance.

R/release-controls.json passes a failing-before control and seven real Chromium 151 cases: normal, delayed down ACK, rejected down after send, pending down ACK, earlier deadline release, rejected up after send, and pending up ACK. The old resolved branch produced no release at 130 ms with down pending; it released only after down resolved. The repaired branch produced exactly one trusted DOM down (buttons=1) and up (buttons=0) in each injected case. The delayed ACK case completed up before down ACK; the deadline case released at 20 ms. The rejected/pending cases remained UNKNOWN. These failures were deliberately injected around real transport sends; no natural Chromium transport stall is claimed. R/adapter-controls.json additionally passes five resolved-adapter integration cases with explicit transport/identity/clock doubles, including UNKNOWN and nonrecoverability after a post-send BeforeInputRefusal. Those five cases are not browser evidence.

The production web-input-probe, guest Worker and render Worker were then used with a generated native control fixture, without launching the original game:

The successful target control above used the shared checkout frontend. Eight files had advanced beyond retained runtime commit d14a8e882f9cee83408ba67587591f90bd921957. Their exact retained bytes were verified recoverable from Git in R/retained-git-recovery.json. R/runtime-http.cjs serves the seven changed frontend JS files directly from those verified Git bytes in memory, without changing shared files or archiving sources. The eighth is a non-browser callback census. control-retained-final/ actually launched both production Workers through this server with the same retained module, completed guarded acquisition and observer arm, and preserved screenshots and identities. It did not complete the relative-motion/target-click sequence. The immutable local window was 11:06:44.606–11:31:44.606Z; its normal quit ran at 11:25:44, browser exited 11:25:46.633Z (code 0), driver code 0. The terminal capture's “single generated control session required” error follows app closure and is not an input ACK or world-action result. A separately named immutable local window was created for a fresh control; the disk guard refused it before its attempt marker/browser. No deadline was extended on an active browser.

Earlier setup failures are retained: the first adapter wrapper could not find its exact source anchor and refused before launch; another full-shell test found an observer helper resolving the predecessor's expired gate. The wrapper was corrected to compile that helper in this run, and the subsequent target control passed. None of these runs replayed UNKNOWN input. Expired captures were refused; subsequent observations were newly captured and reviewed.

The original licence authorization, exact 1,921-original recovery manifest and 2,274-entry retained runtime manifest are preserved unchanged as metadata in R/preserved-inputs.json. The proposed R/runtime-pins.json adds this run's repair helpers and marks qualification incomplete; it has not been staged remotely. Provisioning/retrieval/original-run helpers are preparations, not executed box lifecycle evidence. Original media remains referenced by existing path/hash; no source/media archive or media copy was made. The restricted base Dungeon Siege lane was untouched. There is no FPS, audio or parity claim.

R/cleanup-local.json verifies all 72 recorded runtime PIDs absent, 12 ports closed, six profiles absent and this run's messageboard watcher absent. All local browsers exited normally. Box ID/state/expiry and remote archive checks are not applicable: allocation count is zero, no archived box was resumed and no remote shutdown was performed. There is no box custody to hand back.

Recovery uses Git plus R/owned-harness.patch (65,284 bytes; SHA-256 aa36da9cb148f681761b3d2fdc8d588e953dd9e4802971fa2c4756b01d732a58), which contains only this run's new helper hunks. The initial checkout SHA and relevant tracked diff are retained separately. All 33 run CJS files pass node --check. No production source was changed; the publication is this note, so no production build is claimed. Evidence inventory and publication are recorded before R/result.json is written last.

Next: restore sustained free space above 2 GB, finish the exact pinned full path qualification in actual Chromium before allocation, then use one fresh UUID-idempotent box with an immutable browser deadline below 30 minutes and 40 minutes including retrieval/normal shutdown. Only fresh reviewed captures within 15 seconds can authorize ordinary original-world input. The release repair and fixture counters alone do not prove player action.