refactor: keep test-only code out of production exports

[thymikee]

Summary

check:production-exports listed 59 production-unused exports but never failed: fallow 2.x --fail-on-issues ignored the gate's warn severity. This clears them before the fallow 3.32 upgrade (next PR in the stack) makes the gate enforce again, as #1282 intended.

None has an external consumer. GitHub code search across the 19 public repos importing agent-device (including tester-army/e2e / @e2edev/mobile) found no use, none is documented, and none is reachable from src/sdk/* or src/ai-sdk.

• Test support leaves production modules: the maestro conformance projection joins its harness in packages/maestro/test/conformance/; helpers move to co-located *.fixtures.ts.
• Tests import source modules instead of facade re-exports only tests read, or drive the production caller (buildIosInteractiveSnapshotPresentation, iosSnapshotComparisonIdentityKey, resolveElementReportedTwice). Test-only wrappers and a duplicate equality predicate are removed.
• chore(gates) records seams fallow cannot see. Among them are the staged managed-local lease foundations (ADR 0021). Their Simlock side, simlock#70, was closed as not planned. Removing that feature changes persisted record formats, so it is left for a separate decision.

80 files. The moved conformance projection gets new health-baseline entries for its pre-existing complexity findings: fallow 2.104 counts a moved file as changed, and the old path never needed an entry. An unused test helper is also dropped.

Validation

Tested 600c3cb39 (rebased on main 19203c620): pnpm check:affected --run passed (full check set, selected fail-open by the tooling and config changes). maestro:conformance 62/62 and scripts/ios-snapshot-differential.test.ts 12/12 (node:test suites outside vitest). Pure tooling and test refactor, so no device run applies.

🤖 Generated with Claude Code

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.16 MB	5.16 MB	-93 B
Package (unpacked)	5.16 MB	5.16 MB	-93 B
Package (download)	1.55 MB	1.55 MB	-44 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	27.7 ms	26.2 ms	-1.5 ms
CLI --help	82.7 ms	81.3 ms	-1.5 ms

[cubic-dev-ai]

All reported issues were addressed across 80 files

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>

[thymikee]

The code looks good at 9afc83d. I have one design question below, so I am not marking it ready yet. Smoke Tests was still running when I checked. The diff only removes exports that tests read and moves test helpers. The daemon, Apple, Android and CLI code bodies on the smoke route are unchanged, so I found no overlap with it. I did not run fallow, so I could not confirm the health baseline state or that the gate is clean at head. No conflicts. Smoke Tests must finish green on 9afc83d before merge.

Not blocking, and you can take or leave these. The PR body says the moved conformance-normalize keeps its baseline counts, but at the merge base there is no baseline entry for the old path, so the two complexity_critical and two crap_critical entries at fallow-baselines/health.json:533 are new. Please say so in the body, or drop the entry if the move did not cause them. The new ignoreExports entries at .fallowrc.json:371 keep test-only re-exports (SCRIPT_FLAG_COMMANDS and scriptFlagEntries, acquireDurableCaptureRecoveryAuthorityBeforeDeadline, collectedRecordingPath in stop-sequence.ts). The rest of the PR points tests at the owning module, so these tests could import internal/script-utils.ts, recovery-authority.ts and artifact-paths.ts, and the re-exports and entries could go. The fixture selectMaestroPositionMatches at packages/maestro/src/internal/runtime-target-ranking.fixtures.ts:8 shares its name with a different function in runtime-target-position.ts:17, so a name like resolveMaestroPositionForTest would avoid confusion. The tests at packages/selectors/src/interaction-targeting.test.ts:458 are titled "through the rule on its own" but now call the composite resolveElementReportedTwice, so they need a new title.

Could the enforcement footprint be smaller? Production net is -48 lines, but the PR adds a test-only subpath (capture-kit/audio-probe-descriptor-fixtures) and many ownership entries. If tests imported owning modules for every symbol that only tests read, only the fixture subpath would stay, which follows the existing replay-port and selectors fixtures pattern. The managed-local allowlist entry is fine as a recorded seam until there is a decision to delete the staged managed-local lease foundations, since their owner kind is persisted in device-claim and durable-resource records.

Both cubic-dev-ai P2 threads are fixed at this commit and can be resolved: #3369 (comment) (shellQuoteIfNeeded now quotes casePath) and #3369 (comment) (DifferentialCase is no longer exported).

[thymikee]

Thanks. I tried the smaller footprint, and two gates explain why it stops where it does.

Design question (enforcement footprint).

• Importing the owning module directly: R11 refuses a relative import into a package from outside it ("import the package specifier instead, or dual specifier/relative loads instantiate the module twice"). A root or cross-package test can only reach a module through a package subpath.
• Routing through an existing fixtures subpath: I tried this for collectedRecordingPath, re-exporting it from @agent-device/capture-kit/recording-artifact-fixtures. The eager-closure budget then failed (scripts/__tests__/eager-closure-budgets.test.ts). The fixture entry would evaluate artifact-paths.ts, which it didn't at the merge base, and the gate refuses any new eager edge on an entry module. So it stays a re-export from recording-stop-sequence, beside nativeRecordingPath, which production reads through that same subpath.
• A new fixture subpath per symbol: this would add a manifest entry plus a layering pin each. That's more footprint than the one exemption entry each of SCRIPT_FLAG_COMMANDS/scriptFlagEntries (two names in the existing "test-tree consumers" entry), acquireDurableCaptureRecoveryAuthorityBeforeDeadline and collectedRecordingPath carries now. Neither ad-script nor durable-capture has a fixture subpath.

The other new entries are cases fallow can't see: the provider plugin default export, the lazy import() member, and the members the contracts facade-exhaustiveness rule requires. The managed-local entry is removed again in #3372, which deletes the feature.

Non-blocking items, all taken:

• The body and the gates commit now say the conformance-normalize health-baseline entries are new: fallow 2.104 counts a moved file as changed, and the old path never needed an entry.
• The fixture is renamed resolveMaestroPositionForTest.
• The two wrapper-control tests are retitled "… through resolveElementReportedTwice".

Head: 600c3cb39. pnpm check:affected --run passed on all three stacked heads.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-10 17:18 UTC