Authenticated ordinary launcher-to-child observation

Evidence: scratch/runs/20261117-neverhood-authentic-child-handoff-room/.

test/run.js now supports a one-shot child observation capability through lib/cli-child-observation.js. The controller arms it against a verified full native review before Play. It binds the actual input owner, epoch, identity, nonce, parent USER actor, expected installed executable path and SHA-256, module, runner, bridge and sealed source inventory. Ordinary VFS CreateProcess records the real OS fork and PID start identity synchronously. Unrelated, modified, repeated, late or wrong-actor launches fail the observation capability without replacing the guest's ordinary process behavior.

The child validates its actual parent, loaded executable, module and source identity. It publishes its own readiness only when its listener is listening. The controller checks the child PID/start token/boot ID and listener inode ownership before requesting authenticated child USER readiness. The child's ordinary actor must pass the existing native readiness and complete capture predicates. Launcher exit must independently be zero. A parent receipt never stands in for the child's scene or input owner.

After the armed Play, the controller sends no launcher snapshot. A retired launcher admits only the already prearmed independent physical release until its original 1750 ms deadline, then closes its endpoint. This does not execute more guest code. Child observation and capture must finish within Play's original 3000 ms deadline. Child input gets a distinct, permanently bound child controller after a new personal image review. An UNKNOWN latch prevents all subsequent transport, including cleanup, child adoption and rebinding.

Child native Quit requires the child's own guest exit record after independent UP, its actual OS wait, and both pipe EOFs. Host cancellation and runner errors publish a failure marker even if a guest zero-exit record already exists. Host cleanup therefore cannot qualify as Quit. The existing 500/1750/3000 ms deadlines and 15/30 second image freshness, focus, modal, generation, actor, target and nonce predicates remain in force.

The authored fixture uses a real native dialog button, ordinary CreateProcess, launcher ExitProcess(0), and a surviving child with two native dialogs. Its controller authenticates the child's own module/source/process/USER identity, captures the public child scene, consumes child input with independent UP, and proves native child Quit zero with actual waits and both EOFs. Negative tests cover foreign/unrelated children, source mismatch, dead/reused PIDs, wrong actor/owner/nonce, late transitions, parent failure, child cancellation and permanent UNKNOWN with zero later transport. The new observation path prevents the former parent-exit/in-flight-snapshot race by never issuing that snapshot after the prearmed Play.

The preparation continues to select the qualified published module and CLI components through exact Git plus explicit owned binary patches. Foreign guest-thread-host, NE bridge, CLI Edit and GL working changes are excluded. No source archives, media bundles, frame series, boxes, workers or public deployment are used.

Qualification uses the retained published module and exact sealed component inventory. The new runner, bridge and controller are exercised by real native handoff, child input/UP/Quit, parent failure, child cancellation, inherited ISO and permanent refusal cases. Unchanged historical core libraries are reused only after byte comparison; the qualification explicitly does not claim a fresh full historical native-suite pass. Earlier broad fixture failures and a later three-second deadline refusal remain in the evidence, without pass credit. Concurrent parent/child failure publication preserves the first terminal record and removes its temporary file; eight real OS publishers exercise that race.

Published implementation: 95990aee3. Its locked build and complete sealed qualification passed before the single authorized authentic installed-shortcut run. That run restored the authentic 65-record/58-file state, preserved WININIT, opened SETUP95 through the generated ShellLink, and consumed the personally reviewed Play button. It forked the expected installed NHC with authenticated prearmed provenance and recorded launcher guest ExitProcess(0).

The original stopped before child observation: independent UP encountered the controller's old fixed 200 ms HTTP timeout during synchronous child VFS export. The actual parent fork and guest zero exit occurred before the unchanged 1750 ms UP deadline, but the premature transport timeout had already failed the release. The permanent UNKNOWN latch closed all later transport; host cancellation and actual waits/EOFs are cleanup, not native Quit. No child scene, room action or authentic save was earned, and no second original is authorized.

The subsequent repair gives the one prearmed release HTTP request only the remaining time before its original absolute UP deadline. It does not change 500/1750/3000 ms limits, image freshness or input/context predicates. Real TCP negative tests deliberately disconnect after 300 ms and stall past the absolute deadline: the first must reach the actual disconnect instead of the old 200 ms timeout, the second must time out at the deadline. Both record no ACK and one request without retry; no synthetic successful transport is used. Real native handoff/input/UP/Quit and terminal failure regressions are rerun separately.

Final repaired-source qualification passed in fixture-release-final: native handoff, child game input/UP/Quit, parent Exit23, child cancellation, source and identity refusals, inherited ISO, installed provenance, controller units, desktop participants and all twelve ordinary native controller cases. The locked build passed again. An earlier repaired-source authored Quit run lost its UP connection before child shutdown under load; it remains a recorded failure, without pass credit or a deadline/fixture clock change. A fresh authored run of the same sources passed. The authentic original was never repeated.