refactor(daemon): package-local runtime fakes for the remaining daemon ports [11a-4]

[thymikee]

What

Daemon-extraction slice 11a-4: eliminates every edge where a file under src/daemon/ imports (or vi.mocks) a non-package module outside src/daemon/. Residue scan: 0 edges (node scripts/daemon-edge-residue-scan.mjs, committed in this PR; definition below).

Stacked on refactor/daemon-test-data-fakes (#3517).

How

Daemon suites now answer platform questions at the daemon's own ports, and every port member is a plain vi.fn spy with a canned default answer that carries no platform rule (owner decision on review):

• platform-services port (src/daemon/__tests__/platform-services-fixture.ts) — all six members (appleSessionObservation, deviceBootObservation, ensureLocalDeviceReady, resolveRequestedOpenSurface, validateOpenRelaunchTarget, resolveSessionAppBundleIdForTarget) are spies; the former near-verbatim copies of platform-runtime-open-target.ts, platform-runtime-device-boot.ts and platform-runtime-device-ready.ts are deleted, and so is the shared fixtureResolveSessionAppBundleIdForTarget the lifecycle fixture imported.
• application-lifecycle runtime binding — resolveOpenTarget, prepareApplicationOpen and openApplication are spies whose canned answers delegate to the contracts' canonical direct-lifecycle members (identity pass-through, readiness pause, admitted-hint apply). The macOS frontmost-surface rule and the android stopped-AVD discovery branch are gone. The real composed fixture lives root-side (src/__tests__/daemon-composition/application-lifecycle-runtime-fixture.ts) and is built on the same seam set, so both lanes observe one spy family; the gateway and session-command harness take an injectable lifecycle-operations loader.
• runtime-hints — hint apply/clear are observed at the fixture's port spies; assertion bodies unchanged.
• root-module mocks — the four replay-repair platform-runtime-runtime-hints.ts mocks and the session-open-runtime platform-runtime-apple-application-tools.ts mock were dead (no daemon file loads either module; hint apply/clear reach the device through the admitted binding's operations) and are deleted, not retargeted.

Tests that assert what the REAL composed rules answer do not stay in the daemon lane: they move to the composition lane (src/__tests__/daemon-composition/), where the lane's own fixture hands them the root-composed services and the real composed lifecycle. Assertion bodies are copied verbatim; no assertion was edited to fit a fake. The device-ready suite (it pins what the composed readiness port does per device family) whole-file-moves to the composition lane; 19 composition-lane suites that imported the daemon's services fixture now import the composition lane's own real one.

Rule-ownership relocation (this revision, per review)

Sorted with the throwing-fake / canned-success probes against the spy fixture:

• Relocated to the composition lane (18 tests, names unchanged): session-open-existing open-target/surface cases ×8 → session-lifecycle/session-open-existing.test.ts; open --relaunch target-validation cases ×5 → handlers/session-relaunch-guards.test.ts; device-ready TTL-plus-real-readiness cases ×5 → device/device-ready.test.ts.
• Kept in the daemon lane, retargeted to the port spies (assertion bodies unchanged): boot-probe answers in request-router-open / request-router-open-claim ×3 now arm observeBootTimeMsFixture; open-target identity ×2 (session-open-runtime, session-open-execution-runtime) answer resolveOpenTargetFixture per case; the app-event lifetime case answers resolveSessionAppBundleIdForTargetFixture; the claim-rollback case injects its failure at prepareApplicationOpenFixture.
• Dead mocks deleted: 5 (fix for the five remaining root-module mocks).

Per-file test-name parity vs the previous head holds: 2671 names, identical multiset; 18 relocated with their names, 0 renamed, 0 added, 0 dropped.

Residue scan (definition, as requested)

scripts/daemon-edge-residue-scan.mjs walks every .ts/.tsx/.mjs under src/daemon/ and matches each specifier form independently (the earlier ad-hoc scan used one alternation bridging import|export … [\s\S]*? from, which swallowed a vi.mock() specifier sitting between an earlier importOriginal<typeof import(...)> and the file's next static from — that is how five mocks hid under its "0"):

1. static import/export … from with the keyword anchored at a statement start (line start or ;), so no match ever spans statements (covers import type, multi-line specifier clauses);
2. side-effect import '…';
3. dynamic import('…') and require('…') with string or no-substitution template-literal arguments (this also catches importOriginal<typeof import('…')>);
4. vi.mock / vi.doMock / vi.unmock / vi.importMock / vi.importActual / vi.importOriginal first argument, any position in the file, string or template literal.

A relative specifier that normalizes outside src/daemon/ is an edge; bare specifiers (workspace packages, externals) are allowed. Exit 1 on any edge. Verified: 5 edges on 855bd0f, 0 on this head. Once the daemon is its own package, 11b's R11 package-boundary rule (#3546) enforces this permanently in check:layering; until then this committed scan is the measurement.

Gates

• unit-core lane (vitest run --project unit-core, full): 1591 files / 13447 passed, 1 skipped, 0 failures; within it daemon lane 316 files / 2306 tests, composition lane 97 files / 638 tests, both 0 failures
• check:quick, check:layering, check:di-seams, check:gate-manifest, check:fallow, check:affected --run, eager-closure-budgets + test-file-size ratchets, pnpm format: OK

Pure moves / renames dominate the diff; hand-written additions are the port fixtures, the install seam, the composition siblings, and the committed scan.

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.18 MB	5.18 MB	0 B
Package (unpacked)	5.18 MB	5.18 MB	0 B
Package (download)	1.56 MB	1.56 MB	-1 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.3 ms	27.2 ms	+0.9 ms
CLI --help	78.3 ms	79.8 ms	+1.5 ms

[thymikee]

Coordinator: rebased onto the current tip of the PR below (stack sync, no content changes).

[thymikee]

The daemon-side test fixtures duplicate root platform rules, and five root-module mocks remain in src/daemon, so this is not ready at 855bd0f. All 14 checks pass, and there are no conflicts.

daemonPlatformServicesFixture in platform-services-fixture.ts re-implements root composition code. fixtureEnsureLocalDeviceReady copies ensureLocalPlatformDeviceReady, fixtureDeviceBootObservation copies platform-runtime-device-boot.ts, and the open-target helpers (fixtureResolveRequestedOpenSurface, fixtureValidateOpenRelaunchTarget, fixtureResolveSessionAppBundleIdForTarget) are near-verbatim copies of platform-runtime-open-target.ts. The copies are not spies, and nothing pins them to production. The Linux and macOS surface rules were already merged into resolveSurfaceFor, and appleSessionObservation became a no-op. So a change to the root open, relaunch or readiness rules can leave the daemon suites green with stale behavior. The rule should be that each platform rule has one implementation and every consumer imports it. Please either move the open-target, surface, readiness and boot rules into the owning package leaves so root and the fixture both import them, or make the services members plain vi.fn spies with canned answers and keep the rule-pinning cases in the composition lane (session-open-surface.test.ts already exists). The lifecycle fixture's resolveOpenTarget, prepareApplicationOpen and hint reapplication need the same treatment.

Issue 11a-4 requires that no src/daemon file imports or vi.mocks a non-package module outside src/daemon, and the PR body reports 0 edges. Five files still do. Four replay-repair tests mock platform-runtime-runtime-hints.ts (transaction.test.ts:45, and the record-exclusion, close-ordering and force variants), and session-open-runtime.test.ts:21 mocks platform-runtime-apple-application-tools.ts. Nothing under src/daemon imports those modules now, so these look like dead mocks. Please delete them, or retarget them to the fixture spies (applyRuntimeHintsFixture, clearRuntimeHintsFixture), then re-run the scan, confirm zero edges and say how the scan is defined.

Could the services members be spies with canned answers, with the lifecycle-ops and services fixtures as thin vi.fn doubles? Cases that assert real open-target, surface or readiness rules would stay in the composition lane. If the daemon suites truly need the real rules, extract them once into package leaves that both src/platform-runtime-*.ts and the fixture import. The first step is to decide whether the rule owner is the package leaves or the composition lane. Either way, the daemon-side twin should go before this slice lands.

I did not run the test lanes, so the 96-file, 624-test composition run and the daemon lane results are the author's. I did not execute the five mocks to confirm they are dead. I also could not find the residue scan in the repo, so the count above comes from my own reading of the import specifiers. Before merge, the five mocks need to go and the duplicated rules need one owner.

[thymikee]

@thymikee all three findings addressed on top of 855bd0f (new commits only, 855bd0f..2b85844; #3546/#3520 codemods can re-run on top). Mapping:

1. Duplicated platform rules in daemon fixtures — composition lane owns the rules.

• daemonPlatformServicesFixture members are now all plain vi.fn spies with canned answers: boot probe → {observed:false,reason:'unobserved'}; readiness → resolves true; surface → existingSurface ?? 'app' (no platform branch); relaunch validation → accept; open-target bundle id → pass through currentAppBundleId; observation members → undefined. Every re-implemented rule deleted: fixtureEnsureLocalDeviceReady, fixtureDeviceBootObservation, fixtureResolveRequestedOpenSurface, fixtureValidateOpenRelaunchTarget, fixtureResolveSessionAppBundleIdForTarget (including the shared export the lifecycle fixture imported). Neither fixture file imports any contract rule helper or platform package any more.
• Lifecycle fixture: resolveOpenTarget, prepareApplicationOpen, openApplication are spies whose defaults delegate to the contract's canonical direct-lifecycle members (identity pass-through / readiness pause / admitted-hint apply). The macOS frontmost-surface branch and the android stopped-AVD discovery branch are deleted — those were the real owners' rules.
• Test sort (throwing-fake + canned-success probes; no assertion edited to fit a fake):
  • Moved to src/__tests__/daemon-composition/ (18 tests, names verbatim): session-open-existing open-target/surface cases ×8 → session-lifecycle/session-open-existing.test.ts (real composed services + real composed lifecycle); open --relaunch URL/Android-binary validation ×5 → handlers/session-relaunch-guards.test.ts; device-ready (pins composed readiness per family) whole-file ×5 → device/device-ready.test.ts.
  • Kept daemon-side by retargeting the observation handle (11 cases, assertion bodies unchanged): boot answers → observeBootTimeMsFixture (request-router-open, request-router-open-claim ×3); open-target identity per case → resolveOpenTargetFixture (session-open-runtime, session-open-execution-runtime ×2); app-event lifetime → resolveSessionAppBundleIdForTargetFixture; claim-rollback failure injection → prepareApplicationOpenFixture.
  • The 19 composition-lane suites that had been importing the daemon's services fixture now import the lane's own root-composed one (src/__tests__/daemon-composition/platform-services-fixture.ts).

2. Five remaining root-module mocks — executed, they were dead; deleted.
Nothing under src/daemon loads platform-runtime-runtime-hints.ts or platform-runtime-apple-application-tools.ts any more (hint apply/clear reach the device only through the admitted binding's operations, observed at applyRuntimeHintsFixture/clearRuntimeHintsFixture; the prewarm spy has no caller). The four replay-repair mocks and the session-open-runtime mock are gone, suite still green. Bonus: request-router-open.test.ts was still mocking platform-apple/simulator-boot for a leaf the daemon doesn't load; retargeted to the boot spy.

3. Why the scan said 0 — fixed, and the scan is now in-repo.
The old regex was one alternation: (?:import|export)[\s\S]*?from\s*['"]…['"] | import(...) | vi.mock(...) | importOriginal<typeof import(...)>. Regex alternatives are tried at each position left-to-right, and [\s\S]*? crosses newlines: starting at an earlier import, the first alternative gobbled everything — including the whole vi.mock('../../../platform-runtime-runtime-hints.ts', …) block — up to the file's next from '…' and captured that specifier (an allowed one). The root-mock specifiers were consumed mid-bridge and never reported. I reproduced this: the old scan prints 0 on 855bd0f while the corrected one prints exactly the 5 you listed.
Committed as scripts/daemon-edge-residue-scan.mjs: matches each specifier form independently — statement-anchored static import/export … from (incl. import type, multi-line clauses), side-effect import, dynamic import()/require() with string or no-substitution template args (covers importOriginal<typeof import()>), and vi.mock/doMock/unmock/importMock/importActual/importOriginal first arguments at any position. Relative specifier resolving outside src/daemon = edge; bare specifiers allowed; exit 1 on any edge. Verified 5 edges on 855bd0f, 0 on 2b85844. PR body now carries the definition verbatim, and notes 11b's R11 (#3546) owns this permanently once the daemon is a package.

Lanes/gates. Full unit-core: 1591 files, 13447 passed / 1 skipped / 0 failures. Daemon lane 316 files / 2306 tests and composition lane 97 files / 638 tests: 0 failures. Per-file test-name parity vs 855bd0f: 2671 names identical as a multiset (18 relocated with names, 0 renamed/added/dropped). check:quick, check:layering, check:di-seams, check:gate-manifest, check:fallow (spies kept unexported; scan registered in .fallowrc entries), check:affected --run, eager-closure/size ratchets, pnpm format: all OK.

[thymikee]

Coordinator: rebased onto #3517's new head 2354fd0, which now sits on current main. The rebase applied cleanly and none of this PR's commits changed. Verified at this head: scripts/daemon-edge-residue-scan.mjs reports 0 edges, including the tests main added; check:quick passes; the daemon and composition lanes pass (488 files, 3508 tests).