Alice ordinary dialogue, knife and attack — 2026-11-06

Result: generated supervisor and actual launch/request adapter controls PASS. Original Alice never launched: the static media upload timed out, then the explicitly authorized recovery resume returned403 and was held without retry. No completed dialogue, knife or attack claim. No production edit, build, browser, emulator or module execution occurred. Evidence: scratch/runs/20261106-alice-dialogue-qualified/.

Initially queued behind Lords2 892904 and Aranna935939, this worker remained alive with read-only/tiny owned preparation. Coordinator22:21:01 granted tiny fixtures;22:26:51 granted actual adapter integration and one separate boat; 22:32:37 granted one original only after remote pins/TTL verified. Generated work proceeded under those grants, without touching the shared emulator lane.

Read CLAUDE.md, TODOS NEW-GAME-ALICE-DEMO-20261009, authentic-ground, provisioned-dialogue, observed-timeout, wind-crossing-native, second-gnome-knife, timber-knife-attack, released dialogue-lifetime note and supervisor draft. The draft was NEVER executed. It lacks integrated request admission and cannot authorize an original. The old destructive ProcessDesktop poll remains UNKNOWN; do not retry it, increase its3000ms timeout or infer causal IPC repair from the later diagnostic's success.

Distinct ordinary progression

Personally viewed retained alice-wind-crossing-native/far-landed.png: Alice stands dry on the far platform facing a wooden ramp, with a gnome ahead. Personally viewed retained alice-second-gnome-knife/jamb-right.png: Alice faces the central timber opening and purple-lit passage. These are historical images, not new execution, exact dialogue text or knife/attack evidence.

  1. Under a separate original grant and freshly assigned boat, ordinary balanced Enter selects New Game/Easy. Allow the natural opening. Navigate using reviewed landmarks and short native input segments; historical route totals are guidance, not a blind replay or a timeout reproduction route.
  2. Reach the accepted plank crossing/dry shelf with ordinary W/Shift/Space and camera movement. Crossing and repeated walkway movement earn no new credit. Do not load old held-phase, relocation or guest-memory capture adapters.
  3. Approach the second gnome along the far shelf. Release controls and review the natural dialogue sequence in short neutral segments, at most64 credits each, reducing near transitions. Require the sequence and return to normal player control; elapsed neutral credits alone cannot prove completion. The retained original gvillage_rabbit.scr lines437–455 identify the natural second-gnome sequence as Torchgnome2_Dialog, including the essence-cat sequence and End_Cinematic_Fast; do not invoke or skip these functions.
  4. Follow the visible eastward passage. At a matching timber jamb, a short ordinary right strafe follows the retained central-opening reference. Review the next forward segment to prove clearance; the old failed capture did not.
  5. Continue east, then south only when the scene supports it, toward the script's knife altar. Historical coordinates are navigation context only, never write/teleport targets. Let the natural sequence end and require a controllable Alice visibly holding the knife. A cinematic fakeplayer knife attachment is insufficient.
  6. The existing original readme.txt lines156/158 identify Mouse1 primary and Mouse2 alternate attack. Existing test/run.js parses mousedown:X:Y and mouseup:X:Y as balanced canvas mouse input; verify the pinned runtime's matching implementation and new viewport before use. Capture neutral ownership, one balanced ordinary attack, and a changed swing/pose or target response. Distinguish swing from damage. If evidence cannot safely select an action, stop and retain the boundary.

Elapsed supervisor and request admission to qualify

Preserve CLI waitForControlBatch extending deadlineMs during frozen review. Add only an external immutable monotonic process deadline starting before spawn, including startup, frozen review and pending I/O. Fix ordinary quit, TERM, KILL, terminal observation and final cleanup checkpoints before launch. No extension. Record actual elapsed overrun honestly; OS scheduling is not a hard realtime guarantee. The historical crossing took1770.833s with heavy diagnostics, while the later900s diagnostic reached only4926batches and retired917.637s. A later original grant must explicitly include enough progression/review/recovery time.

Integrate the request writer with the supervisor's shared admission lock and immutable schedule. Check admission, live child identity, retained first error, label uniqueness, command bounds and pending state before each single dispatch. Create the pending claim before network submission; retain it as UNKNOWN after timeout/disconnect, never resubmit or equate cleanup with completion. Only a validated terminal response permits the next dependent request. Close admission before quit. If admission lock is busy or any request is uncertain, withhold the ordinary control quit and use bounded process escalation. Do not let a blocking HTTP operation or admission lock delay escalation. Quit is attempted at most once. Preserve first-error bytes with exclusive creation, with later failures separate. Track both leader exit and descendant cleanup; leader exit alone is insufficient. Record terminal return code, signals, pending outcomes, group absence, deadline compliance and owned helper cleanup independently.

Under the limited fixture grant, tiny generated fixtures cover active progress, frozen idle with an intentionally extending internal active deadline, pending request, child exit0 and exit7, unavailable control, ignored TERM/KILL escalation, orphan child, admission race, duplicate label/output and invalid checkpoint order. Assert no resubmission, no postclosure dispatch, first-error retention, unchanged absolute schedule, bounded exit and truthful UNKNOWN. These tests passed. Generated-control success would qualify only the supervisor, not Alice gameplay.

Owned admission.py, supervisor.py and launch-gate.py began as UNEXECUTED preparation drafts and were tested only after the explicit limited grant. They integrate the shared gate, pending custody, separate quit helper, Linux descendant reaping and five fixed checkpoints. The command-line entry accepts only owned generated fixtures. The old released predecessor draft was never executed. The new helpers are qualified for the generated cases below; they admit no original by themselves.

generated-controls-01.json records all10 initial cases PASS. The final generated-controls-02.json records all10 cases PASS on the second helper revision, seven illegal payload rejections, ten exclusive duplicate-output refusals and invalid checkpoint-order refusal, with all21 launcher PIDs recorded/reaped. generated-admission-extra-01.json adds two validated live requests, duplicate label refusal, postclosure refusal before claim and eight concurrent secondary errors preserving exact primary bytes. fixture-childidentity-01/terminal.json proves an exited/zombie child cannot admit a request before its leader is reaped.

Active and frozen-idle fixtures exited0 by one ordinary quit before escalation; the idle child's internal active deadline extended while the external schedule did not. Pending and occupied-lock cases withheld quit and retained UNKNOWN. Unavailable control attempted quit once, retained UNKNOWN and escalated. Exit7 retained its actual exit code; ignored TERM required KILL; an orphan after leader exit was signaled/reaped rather than treating leader exit as cleanup. Expected fixture errors stay in their own fixture directories. Every final fixture's reported elapsed time remained inside its1.1s immutable deadline. No real scheduling-overrun fixture or instantaneous realtime guarantee is claimed.

The first cleanup-summary attempt raised KeyError('procStat') on the focused identity fixture's shorter metadata. Root first-error.json retains it. The read-only summary was corrected without rerunning a request. Initial negative CLI launch PIDs were not logged, although subprocess.run waited/reaped them; the final set fixes that missing metadata and retains the first-set uncertainty. generated-cleanup-final.json independently records all known PIDs absent and all generated private ports refused. No fixture processes remain owned.

actual-adapter-controls-01.json then qualifies the actual original-adapter.py and request-control.py entry points with generated active, idle, pending and exit7 children. This covers the final supervisor's launch environment and adapter authorization gate. The original's fixed8123 address belongs to the unexecuted original branch. CLI duplicate and concurrent requests are refused; a provisioning-only grant cannot create an original run. All four generated process/helper groups and recorded launchers were reaped. This does not execute the original CLI/emulator or qualify Alice gameplay. The final read-only generated-cleanup-final-02.json observes80 recorded PIDs absent and18 private endpoint observations refused, including the adapter cases.

Provisioning failure and lifecycle boundary

selected-recovery-verification.json verifies every one of385 tracked source blobs from selected Git277f9a32, all43 existing media/dependency files plus the manifest and both original065267ce/e7c59d3d modules by full size/SHA. These are verified retained bytes; no module was compiled or instantiated. Git and small owned patches are the source archive. Direct provisioning created no source or media archive and copied no media into evidence.

After reading sandbox SKILL.md, the worker persisted/fsynced a UUID in boat-create-key.json before the sole create POST. Fresh bx_3xdr3tsx was created22:31:33.576Z, TTL3600, original expiry23:31:33.591Z; shutdown was delegated. The single80,849,066-byte pak0.pk3 signed PUT timed out during SSL write after the90s transport bound. provision-uncertain-upload.json preserves UNKNOWN. The other438 transfers, including the pin manifest itself, completed. A separate read-only observation found the remote pak absent. Final read-only audit verified 437 manifest entries, zero mismatches, only that pak missing, no unexpected unique files, no native directory or original launch claim and control8123 closed. Unique setup receipts/audit were retained locally before delegated normal stop.

At22:35:13 the coordinator clarified that recovering a confirmed-absent static media upload is allowed and differs from retrying an uncertain guest command. The worker's blanket no-retry interpretation was too strict. The stop attempt had already been persisted at22:35:13.204836Z and requested22:35:13.379Z before that steering was read. Actual normal stop stop_4f6622369449c4e5 completed 22:35:14.144Z, forced=false/error=null. GET confirmed API state archived, stoppedAt22:35:14.126Z and active expiry null. The timing/state mismatch was reported immediately; no second box or automatic resume was attempted.

Coordinator22:36:26 explicitly authorized ONE normal resume of this same box, which had no preceding403. boat-resume-attempt.json persisted its UUID and body at22:36:58.980386Z before the call. POST /sandboxes/bx_3xdr3tsx/resume returnedHTTP403 at22:36:59.063180Z. boat-api-failure-b6aae9c5-98e4-4397-8bf7-a8a40437c300.json records the exact method/route/status/time. No retry, rephrase, alternate CLI, delegation, further lifecycle request or new box followed. The streaming recovery helper was prepared but NEVER executed. The last confirmed box state remains archived/normal-stopped; there is no successful resume or new expiry. This403 is the concrete blocker.

The original grant fixed quit/TERM/KILL/terminal/final at2100/2130/2140/2160/2190 elapsed seconds including review, with180s extra TTL for retrieval/normal stop. Those original checkpoints were never started: complete remote media pins could not be established. There were zero original guest requests, inputs or captures. The historical ProcessDesktop3000ms timeout/outcome remain UNKNOWN and unchanged.

Root first-error.json retains the earlier cleanup-summary metadata error; secondary records retain a source-metadata key correction and later provisioning/ lifecycle failures. The original-source key is testedRuntimeCommit, not commit; correcting that read allowed full local verification. No failure bytes were replaced. No giant archive, repeat historical diagnostic or source fix.

Retained exact recovery references

Accepted crossing runtime: Git 277f9a32c984562ef33b2d32a5c13df949b5f87a. Canonical2125777bytes SHA256 065267ce7c208cb369b1fa930425627b3a238c9e184e5aa46d8ef16143b8b5b8; compat2128390bytes e7c59d3d39d937f5bef162b9f052259075ff1040d49f5b1be40e8534c5e9fdfe. Changed diagnostic Git 625be68f9c2b685797a81f837992f4421d469f16 includes observation458625894. Its retained alice-observed-timeout modules are wine-assembly.wasm.gz (2143540 decompressedbytes, cffca869b541e6eff12ae3a6b74c4476977e0a7e23cd43d8ec5ba8faecd4bcee) and wine-assembly.compat.wasm.gz (2146153bytes, 580f49f3c78d26e6855efa0beb37046d03ea24690ff98cfcb3e94ce5e6cb1ee3). The selected277f9a32 modules were newly verified as retained bytes, never loaded. The changed diagnostic module identities remain prior references only.

That run's source-identity.json.gz, baseline-module-provenance.json.gz, recovery-verification.json.gz, original43-path-sha-references.json.gz and route-with-captures.json.gz identify source closure, tested module recovery, 43 existing media/dependency paths/full hashes and ordinary route payloads. Use Git blobs plus relevant binary diffs and owned new-file patches; no source snapshot or giant archive. Verify exact existing media paths/full SHA before provisioning; reference media instead of copying it into evidence. Explicitly select and verify the actual runtime/module under its later grant.

Crossing and second-gnome RECOVERY.md retain old archive recovery chains: 34e33a0a97a977eb5867b66366e031e0b408c350582b0d0fdacc8584a6e91046 and 9794e67a79ee6db80e7726b7a051f8de14d60fcf034d3912269cff1107c44009. Reuse exact unique references; no full-tree restoration or new archive implied.

Custody returns to the coordinator: bx_3xdr3tsx last confirmed archived after completed normal stop, active expiry null, later resume403 held. All fixture processes/ports were cleaned; local free disk remained above2GB. No shared runtime/source/build ownership was taken. Publish result.json LAST and release owned run/note after documentation publication. Further work needs permitted access/provisioning disposition; do not retry this403 box or treat fixture success as dialogue/knife/attack completion.