[Bug][Priority II] Close remaining dev-up fail-closed availability races

[Chris0Jeky]

Follow-up from the one-time Codex review triage on PR #1896. These are valid nonblocking source-launcher hardening gaps; #1896 remains mergeable because each is bounded availability, narrow race, or test-reliability risk rather than a demonstrated wrong-instance mutation path.

Acceptance criteria

• Bash treats a verified zombie launcher root as exited so Stop does not wait, KILL, or retain state solely for an unreaped process whose ports are already released.
• Stale operation-lock reclamation is atomically bound to the inspected owner/directory instance; a third launcher cannot acquire and then have its lock removed by the reclaimer.
• Both launchers set child-only DOTNET_ENVIRONMENT=Development as well as ASPNETCORE_ENVIRONMENT=Development, so ambient Staging/Production cannot suppress the run-identity response.
• PowerShell termination stays identity-bound through the actual process-tree signal; a root exit and PID reuse between validation and taskkill cannot target a replacement process.
• Multi-listener launcher tests reserve distinct ports through bind, preventing ephemeral-port reuse from manufacturing EADDRINUSE failures.
• Focused Ubuntu/Windows launcher regressions cover each corrected seam without weakening fail-closed state retention or no-signal behavior.

Source review comments

• Bind source launchers to the launched API instance #1896 (comment)
• Bind source launchers to the launched API instance #1896 (comment)
• Bind source launchers to the launched API instance #1896 (comment)
• Bind source launchers to the launched API instance #1896 (comment)
• Bind source launchers to the launched API instance #1896 (comment)

Refs #1875. This issue does not reopen #1895 or the packaged/manual #1876 gates.

[Chris0Jeky]

Occurrence 2026-08-22: PR #1959 head 2781d88, Frontend Unit (windows-latest) red on scripts/ci/dev-up.test.mjs 'reset seed forwards --reset again after a clean stop' — recorded api PID tree did not exit cleanly under taskkill (child-process termination race), FATAL 'Stack cleanup was incomplete'. The PR's diff touches no script file (frontend views/locales/ADR only), so this is the dev-up cleanup race class, not the change. Job rerun requested; if it re-reds on the same assertion the isolated-rerun protocol applies.

[Chris0Jeky]

Second occurrence 2026-08-22 (same wave): PR #1976 head, Frontend Unit (windows-latest) red on dev-up.test.mjs process-cleanup tests ('invalid Vite outcome stderr-marker cleans both trees', 'success uses exact marker URL, isolated env, versioned identity state, and safe Stop') — 'failed startup retained state despite complete cleanup'. Diff touches Today view/locales/e2e strings only, no script files. Rerun requested; two distinct dev-up cleanup assertions have now flaked in one afternoon on windows-latest.

[Chris0Jeky]

THIRD occurrence in one day (2026-08-22/23): PR #1994 — a DOCS-ONLY diff — red on dev-up.test.mjs 'reset seed forwards --reset again after a clean stop' (windows-latest), the same assertion as occurrence 1. Three distinct heads, two of them touching no script: the flake rate on windows-latest now taxes every PR in a busy wave (each occurrence costs a rerun + a human/agent triage). Escalating from 'occurrence log' to 'worth prioritizing': the recorded process-tree/taskkill race in the stop path is the common thread.

[Chris0Jeky]

Occurrence 2026-08-28: PR #2164 at docs-only head 69e62f9dc4603b5e42dc76560d4a6be88c1da631 hit this existing Windows cleanup/state-retention class.

• Run: https://github.com/Chris0Jeky/Taskdeck/actions/runs/33147855593
• Failed job: https://github.com/Chris0Jeky/Taskdeck/actions/runs/33147855593/job/98772830792
• Exact case: PowerShell HTTP 200 without one exact matching run identity never seeds or succeeds / mismatch
• Result: 81/83 launcher tests passed; the subcase failed after 2,193.5 ms because schema-v1 PID state retained API cmd PID 5724 after the launcher process/ports were already clean.
• Local Windows/Node 24.13.1 exact named re-measure: 4/4 passed in 5.99s; mismatch itself 1.81s.
• Adjacent Parse dead-affordance guards with DOM options #2162 and Carry honest model confidence into review #2160 final heads passed this exact subcase in 1.93s and 2.00s respectively, both 83/83.
• [CI][Windows] Stabilize dev-up reset-seed first-use deadline #2161 is distinct: its failed job passed this subcase and timed out only reset-seed first use.

The current diff touches only docs/STATUS.md and docs/TESTING_GUIDE.md; it does not touch the launcher, test, workflow, or package graph. State retention remained fail-closed and safe, but the required check is red. One bounded failed-job rerun will be requested after the parent workflow completes; the PR remains unmergeable until a fresh exact-head green result.

[Chris0Jeky]

Bounded rerun outcome: workflow attempt 2 / replacement Windows Frontend Unit job https://github.com/Chris0Jeky/Taskdeck/actions/runs/33147855593/job/98775821052 passed on unchanged head 69e62f9dc4603b5e42dc76560d4a6be88c1da631.

The full 83-test launcher suite passed, including the previously red PowerShell mismatch subcase; lint, typecheck, build, coverage, and artifact upload then passed in the same 11m18s job. This clears the isolated PR #2164 occurrence but does not close #1898 or explain the underlying Windows cleanup/state-retention race.