[nplez1]

	Files	Added	Deleted	Net
Test	4	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​23	$\color{#cf222e}{\Huge{\mathbf{−}}}$​1	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​22
Prod	1	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​8	0	$\color{#1a7f37}{\Huge{\mathbf{+}}}$​8

ELI5

Four areas of CI were already red on nplez1/main before this branch existed — three test shards and the packaging census. Every one of them has a distinct, reproducible cause; none is a flaky test. This branch fixes all four, so the next PR on this fork can go green.

What Changed

Before: every PR Checks run failed static analysis (from #8 onward), test shards 4/8, 5/8 and 7/8, and the aggregate verify gate — 13 tests in four files. PR #7 was merged with that red CI; the same 13 tests, in the same four files, failed there before this branch existed.

After: all four files pass, locally on macOS and (for the one platform-dependent case) in a run that forces process.platform = 'linux', which is what CI uses.

file	tests	cause	fix
src/main/ai-vault/session-scanner-service-cancellation.test.ts	4	the fork's index-first scan path calls sessionSearch.listSessions(...), but the upstream test's hand-rolled SessionScannerServiceSearch double only has handles() and close(), so every request threw and settled before the cancel arrived	the double answers null, which is its "no index here yet" answer — the scanner fallback the cases are about
src/main/rate-limits/codex-fetcher-probe-shutdown.test.ts	1	since the macOS resolver probe moved off the main thread it spawns /usr/sbin/scutil through the same node:child_process these cases mock, and that spawn lands during a later case's spawn tally	the diagnostic is stubbed to its no-hint path; the hint is orthogonal to probe shutdown and home-lock ownership
src/renderer/src/components/settings/AccountsPane.copilot-credentials.test.tsx	5	the pane mounts CodexLoginLinkNotice, which subscribes via window.api.codexAccounts.onPendingLoginUrlChanged, and the hand-rolled api double only carried codexAccounts.list	the double honours the subscription contract and publishes no link
config/scripts/mobile-web-bundle-packaging-workflow-contract.test.mjs	3	fork-release.yml packs mac and Windows without install-mobile-dependencies, which upstream's packaging sweep added to every packaging job two days after this fork's workflow was written	both jobs install them right after the root install, as every other packaging workflow does, and the census's expected-job list names the two jobs

Nothing here changes product behaviour: three of the four are test-double or assertion drift, and the fourth is a CI workflow that had fallen behind an upstream invariant.

Why

• Why fix the tests and not the product in the first three. Each subject is code this fork deliberately ships: the index-first scan path is local fork work (LOCAL-PATCHES.md names "the indexed list" and the service query/refresh protocol as staying), the off-main-thread resolver probe is upstream's, and the Codex sign-in link notice is upstream's. The tests were the stale side in every case — the mocks and counters had drifted from the interfaces they stand in for.
• Why stub the diagnostic instead of only counting codex spawns. The stray /usr/sbin/scutil spawn also consumes the mock's mockReturnValueOnce queue, so a filtered counter would still hand the wrong child to the wrong probe. Stubbing the module removes the count and the queue hazard.
• Why the workflow got the step rather than an exemption. That census exists so "a new packaging workflow has to be added here deliberately, with its bundle step, rather than slipping in"; an exemption would delete the invariant instead of meeting it.
• Considered and rejected: skipping or deleting the failing cases (each encodes an invariant worth keeping — cancellation ownership, one probe per running home, credential-source presentation, packaging coverage); rewinding fork-release.yml to a pre-sweep shape (it would still pack without the mobile bundle step); building a fuller fake search class than these cases need (the console of it is the null return).

Linked Issue

None to link: issues are disabled on nplez1/orca, and there is no matching upstream issue. These four fixes were requested directly, as follow-up to the CI failures observed on #8.

Visual Proof

N/A — no product UI or behaviour change. Three of the four commits touch only test files and the fourth touches only a CI workflow, so there is nothing to render before or after.

Testing

Ran:

• The four files together: 57 tests pass.
• Widened: src/main/ai-vault, src/main/rate-limits, src/renderer/src/components/settings and the packaging contract — 403 files, 3,461 tests pass.
• Every workflow/CI contract test in config/scripts (this change edits a workflow, and a dozen of those contracts crawl all of .github/workflows): 208 files, 1,885 tests pass.
• The whole static analysis battery CI runs: audit:anti-slop, audit:code-quality:native, check:code-quality:changed against the base commit, check:react-doctor:changed, check:zustand-selector-fanout, check:dead-classes, check:max-lines-ratchet, check:ts-nocheck-ratchet, check:runtime-electron-ratchet, check:reliability-gates.
• pnpm tc — full typecheck across every project.
• Platform check for the rate-limits case, since its failure was platform-dependent: it passes on macOS (where the stray spawn lands in the previous case) and in a copy of the file that forces process.platform = 'linux' (where it landed in the failing case, as it does on CI).

Not run: the full repo-wide pnpm test (sharded across eight CI jobs), the rest of pnpm lint, pnpm build, and the Electron e2e suite. fork-release.yml has never been dispatched from this worktree — the workflow change is validated by the packaging census and by YAML parsing only, not by a release run.

• I manually tested these changes locally
• Automated tests added/updated, or explained why not below

AI Disclosure

Written with the pi coding agent (model: deepseek-flash). Each failure was reproduced locally, its cause traced to a specific commit or interface drift, and the fix re-run on both macOS and a Linux-forced copy before committing.

Review

Agent skill upstream boundary

• Not applicable, or this change follows docs/reference/agent-skill-sharing-upstream-boundary.md and copies or mechanically translates no upstream skill-installer source, tests, fixtures, registry entries, path tables, comments, or documentation.

Notes

• Security: no product code changes. The workflow step runs the existing install-mobile-dependencies composite action with --frozen-lockfile, which pr.yml already runs in four jobs, and its drift guard compares the mobile lockfiles to the workspace files.
• Cross-platform: the rate-limits fix is specifically about platform divergence — it removes a darwin-only spawn from a counter, so the case now behaves the same on Linux CI and macOS. The other three are platform-independent.
• Worth knowing about the workflow fix: install-mobile-dependencies is currently a forward-looking invariant rather than a repair of a failing release. src/mobile-web is still a 2.5 KB placeholder with no mobile imports (pnpm run build:mobile-web succeeds here without mobile/node_modules), so the fork's releases were not failing at the bundle step. The step removes the census failure and aligns the fork's release path with upstream's rule before the bundle grows into it.
• SSH / remote, mobile, persisted state, wire formats: untouched. No protocol, schema or stored value changes, so there is nothing here for an older client or a paired host to disagree with.
• Performance: no product impact. The workflow step adds one frozen install per release job.

Checklist

• This PR is small and focused
• I explained what changed and why (ELI5, the user-facing before/after, the mechanism, and why over the alternatives)
• Before/after screenshots or videos attached for UI changes, or N/A with reason — N/A, no UI change
• Self-reviewed for correctness, security, and performance
• Cross-platform, SSH/remote, and path/shortcut impact considered (or N/A)
• pnpm lint, pnpm typecheck, pnpm test, and pnpm build pass — ran typecheck, the four suites, the widened suites, every CI contract test and the static-analysis battery; full repo-wide test, full lint, build and e2e were not run, see Testing