refactor: remove test-only seams, dead option fields, and duplicate tests

[thymikee]

Summary

Production code reachable only from tests, option fields no caller sets, and tests that restated a sibling:

• DaemonError is now an alias of NormalizedError (field-for-field identical; wire ledger ack in the final gates commit).
• host-kit: one override-path resolver behind the executable and file variants, the canceled-command wrapper inlined, and the abort-aware sleep exported from retry instead of a second module.
• provision-kit drops the downloadTimeoutMs option no caller supplies; Android observation reads blocking dialogs through app-parsers instead of a verbatim copy.
• Deleted shutdownSimulator, providerDeviceAdmission(), CoordinateGesturePayload, getCloseShutdown, prefixItems, and selectMaestroSnapshotMatches: each had zero production references.
• isLikelyStaleSnapshotDrop is imported from capture-kit instead of duplicated; CheckId derives from ALL_CHECKS instead of a second 59-member list.
• Removed tests whose owner remains: a byte-identical plan-validator case, idle-reap and surface-signature cases implied by a tolerance test, a perf positional case subsumed by the case-insensitive one, interactor-catalog subsets of the seven-operation test, hint regexes duplicated by exact pins, a stub-client clipboard case covered through the real CLI, catalog and lockfile tests re-spelling their siblings, an output-economy expectation computed by the helper under test, and a zip fixture nothing imports. The never-run integration-progress-model.test.ts is now wired into check:affected:test.

39 files, net −473 lines.

Validation

Tested at 12b9b97: pnpm check:affected --run passed (format, lint, typecheck, daemon-wire-compat, related unit tests, check-affected and mutation node tests, output-economy project). No device run applies.

🤖 Generated with Claude Code

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-08 13:19 UTC

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.13 MB	5.13 MB	-1.3 kB
Package (unpacked)	5.13 MB	5.13 MB	-1.3 kB
Package (download)	1.54 MB	1.54 MB	-362 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	23.9 ms	22.5 ms	-1.4 ms
CLI --help	65.6 ms	67.1 ms	+1.5 ms

[cubic-dev-ai]

No issues found across 39 files

_{View guided diff | Turn on auto-fix | Re-trigger cubic}

[thymikee]

This PR is ready at 12b9b97. The removed seams, dead option fields and duplicate tests look right to me, and I found nothing that needs a change.

Not blocking: wiring scripts/integration-progress-model.test.ts into check:affected:test in package.json (https://github.com/callstack/agent-device/blob/12b9b97/package.json#L151) is a second change in a deletion PR, and that test may not be selected when only scripts/integration-progress-model.ts changes, so you can register it under the gate that owns integration-progress or explain why affected-selector owns it, or leave it as is.

The affected decision-kernel check failed only because two of eight mutant shards (kernel-errors and selectors-2) were cancelled, so it reported "Incomplete shard set: 6 report(s), expected 8". No test failed. I could not see why they were cancelled. selectors-2 does not touch this diff. For kernel-errors, the changes are a type-only alias and three removed tests whose hint assertions are still pinned per code, so I expect the same result. I did not run any tests myself.

There are no conflicts. Please rerun the two cancelled shards so the check gets all eight reports before merge.

[thymikee]

Addressed at fea10e9: scripts/integration-progress-model.test.ts now runs under test:integration:progress:check, the script the integration-progress gate owns, instead of check:affected:test. pnpm check:affected --run passed at fea10e9; the mutant shards re-run on the push.

[thymikee]

The two cancelled shards are the 30-minute job budget, not a test failure, and they reproduce outside this branch:

• Mutants (kernel-errors) was also cancelled on refactor/ws10-shared-graph-traversals (run 37751781889) this morning. On this branch it reached 180/388 mutants at 15 min after a 9 min 11 s dry run over 1,221 related test files and 9,965 tests; the last two green runs on main (Oct 1, Sep 23) had 1,142 and 1,087 related files with 6–7 min dry runs and finished at 27–29 min, so the shard has outgrown the budget on current runners. It mutates one file, so --shard cannot split it.
• Mutants (selectors-2) carried 1,074 mutants here versus 678 in the Oct 4 weekly sweep; the 2× dry-run time points the same way.

This PR's diff touches packages/kernel/src/errors.ts only as a type alias, and the three removed hint tests are still pinned per code in errors-code-policy.test.ts, so the kernel-errors result would not change. Affected decision-kernel mutants is not a required check (merge state is UNSTABLE, not BLOCKED). The fix belongs in main: either a larger budget for the single-file kernel-errors shard or a test-level split of its related set. I can open that as a separate tooling PR if you want it.

[thymikee]

This PR is ready at fea10e9. The earlier clean review at 12b9b97 still holds, and nothing new needs a change.

Not blocking: deleting selectMaestroSnapshotMatches also removed the only test that resolved a childOf parent selector with its own index, and resolve(childOf).indexed at https://github.com/callstack/agent-device/blob/fea10e9/packages/maestro/src/internal/runtime-selector-resolution.ts#L173 now has no coverage, so changing it to .matches (or the same in the containsChild and containsDescendants branches) would pass every remaining test. You can take or leave moving that case into packages/maestro/src/internal/__tests__/runtime-selector-resolution.test.ts, using createMaestroSnapshotResolver(snapshot).resolve({ text: 'Delete', childOf: { id: 'row', index: 1 } }).indexed with the same four-node fixture, expecting [2].

The failing check looks unrelated to the change. The PR does touch packages/kernel/src/errors.ts and the kernel error tests, so the kernel-errors mutation lane runs. But that lane produced no result at all: it shows 0 killed, 0 survived and 0 total, and only 6 of 8 shard reports were merged. The aggregator fails at the score stage because shards are missing, not because a score is low. Removing tests cannot lengthen that dry run, and the PR does not touch packages/selectors, so the selectors-2 cancellation does not overlap the diff. I did not open run 37751781889 to confirm that kernel-errors was also cancelled on another branch, and I did not run test:integration:progress:check locally. Because no kernel-errors score exists for this head, I could not measure whether the removed duplicate tests change the kill count, though the exact-hint table at errors-code-policy.test.ts:160-193 suggests they do not.

No conflicts. Before merge, a completed kernel-errors mutation shard is needed, either by rerunning it or by raising the job budget. Until then the decision-kernel aggregator keeps failing at the score stage for any PR that touches packages/kernel/src/errors.ts.

[thymikee]

Mutation lane fixed at 70aa643: the only reason the sweep ran here was the one-line cleanup in scripts/mutation/config.test.ts, which LANE_TOOLING treats as a harness change and so selects the canary plus every derived kernel. That line is restored; pnpm mutation:run --list-affected on this head prints [], so the lane selects no shards. The kernel-errors budget problem itself stays a main-side tooling issue (single-file kernel, 9-minute dry run). pnpm check:affected --run passed at 70aa643.

[thymikee]

This PR is ready. The earlier review at fea10e9 found no problems, and 70aa643 only changes one test assertion, so that result still holds. All checks pass, and there is no production code overlap in the new commit. It merges cleanly with current main. I did not re-check the older commits beyond the range-diff, which shows them identical. Not blocking: deleting the selectMaestroSnapshotMatches wrapper also removed the test "nested childOf applies its own index before scoping the candidate", which was the only test that ran the real resolver on a nested childOf with its own index (the code is resolve(childOf).indexed at https://github.com/callstack/agent-device/blob/70aa643/packages/maestro/src/internal/runtime-selector-resolution.ts#L173). You could move it into runtime-selector-resolution.test.ts as createMaestroSnapshotResolver(snapshot).resolve({ text: 'Delete', childOf: { id: 'row', index: 1 } }).indexed, expecting [2], or leave it out.