refactor(test-kit): host the target classification capture fixture

[thymikee]

What

replay-port/target-classification-fixtures (the bottom-tabs real-capture snapshot and the Article-tab target evidence) moves into test-kit as two named exports, and the export is dropped (replay-port exports 31 → 30). replay-port declares test-kit as a devDependency (manifest + tsconfig reference).

Fixture decision

Fixture	Importers by package	Outcome	Reason
bottomTabsRealCaptureFixture, recordArticleEvidence	replay-port 1 (target-classification.test.ts), daemon 2 (session-replay-repair-empty-tail, session-replay-dispatch-selector-miss)	test-kit named exports	Shared across packages. The daemon tests exercise daemon replay repair and dispatch (SessionStore, replay runtime), so they are not replay-port tests and cannot move (case 3, not case 2). The snapshot data uses kernel/contract types only. recordArticleEvidence takes the evidence function as a parameter (recordArticleEvidence(computeTargetEvidence)), so test-kit does not depend on selectors; each of the three consumers imports computeTargetEvidence from @agent-device/selectors/target-evidence itself.
toSnapshotNodes	same three files	merged into the existing buildNodes	toSnapshotNodes only minted e<n> refs; buildNodes is attachRefs, which mints the same refs and additionally sets kind. All three files pass unchanged assertions with it.

Manifest walk (workspace dependencies + devDependencies, transitive): test-kit gains no dependency in this PR. Today its reach is capture-kit, command-registry, contracts, daemon-contracts, host-kit, kernel, plus ad-script and selectors transitively through command-registry and daemon-contracts. With #3686 (drops daemon-contracts) and #3687 (drops command-registry) both landed, its reach is capture-kit, contracts, host-kit and kernel only: it excludes selectors and replay-port, so replay-port → test-kit closes no cycle and selectors can later declare test-kit.

Commits

1. git mv packages/replay-port/src/target-classification.fixtures.ts packages/test-kit/src/target-classification-fixtures.ts (pure).
2. Export from test-kit's entry (sorted position), drop toSnapshotNodes, importer specifiers/identifiers in three files (await recordArticleEvidence() → recordArticleEvidence(computeTargetEvidence), plus the selectors import in the two daemon files), replay-port manifest + tsconfig reference + lockfile, delete the replay-port export.

Test-count parity (vitest run … --reporter=json)

File	origin/main	here
replay-port target-classification.test.ts	19/0	19/0
daemon session-replay-repair-empty-tail.test.ts	6/0	6/0
daemon session-replay-dispatch-selector-miss.test.ts	6/0	6/0

No vi.mock in any of the three. The fixture file itself is not a test; the destination project collects nothing new.

Surface

replay-port exports 31 → 30; no new subpath anywhere; test-kit adds two named exports from its one entry; no barrels or export *. replay-port gains devDependencies: @agent-device/test-kit and the tsconfig reference.

Gates (on the pushed head)

check:quick, check:layering, check:fallow, check:gate-manifest, pnpm format, check:affected --run: clean.

🤖 Generated with Claude Code

[cubic-dev-ai]

No issues found across 10 files

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

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	5.24 MB	5.24 MB	-5.9 kB
Package (unpacked)	5.24 MB	5.24 MB	-5.9 kB
Package (download)	1.58 MB	1.57 MB	-1.6 kB

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.8 ms	26.1 ms	-0.7 ms
CLI --help	77.8 ms	77.7 ms	-0.1 ms

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-10-16 06:01 UTC

[thymikee]

This PR is ready to merge at 99f3a6f. The code looks right, and all 23 checks are green. I did not run the tests or the build myself. The 19/6/6 parity table and the gate results come from the PR body and the green CI. I also did not read attachRefs, so the claim that buildNodes differs from toSnapshotNodes only by setting kind is unchecked, though the unchanged assertions passing in CI support it. I checked for cycles by reading each package.json's workspace dependencies, and I checked tsconfig references only for replay-port and test-kit. There are no conflicts.

Not blocking, and fine to take or leave: repair-hint.test.ts still has a local toSnapshotNodes copy that could use buildNodes from @agent-device/test-kit as a follow-up; recordArticleEvidence takes computeTargetEvidence as a hand-written structural type that could drift, so typeof computeTargetEvidence through a type-only import would pin it if that ever matters; and the private bottomTabsRealCaptureFixture in target-evidence.test.ts can be deduplicated once #3686 and #3687 land.