refactor(replay): move adapter leaves into replay-port

[thymikee]

Summary

Move six dependency-light replay adapters and the paired Maestro observer test into @agent-device/replay-port: script selection, session binding, selector candidate healing, Maestro observation, video recording, and test shard device selection. Daemon callers now import the owning package subpaths. Remove their superseded replay-index reexports and layering declarations.

This dependent #2663 slice builds on #3024. The moved production implementations are unchanged; the observer test uses the package scratch fixture.

Validation

• Exact head 056b65ee1468b0618ec45977e082652009397e54: pnpm check:affected --run passed (610 test files, 5,407 tests; separate 12-test suite passed).
• Focused Vitest: observer, daemon device-selection consumer, and eager-closure budget suite passed (737 tests).
• pnpm typecheck, tracked-file R11/depgraph layering, and repository-wide pnpm format passed.

The remaining daemon replay tree and live .ad/Maestro characterization remain tracked in #2663.

[cubic-dev-ai]

No issues found across 20 files

_{Re-trigger cubic}

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://callstack.github.io/agent-device/pr-preview/pr-3028/
Built to branch gh-pages at 2026-09-28 17:01 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[github-actions]

Size Report

Metric	Base	Current	Diff
Installed (including dependencies)	4.87 MB	4.87 MB	+183 B
Package (unpacked)	4.87 MB	4.87 MB	+183 B
Package (download)	1.46 MB	1.46 MB	+505 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	28.7 ms	28.7 ms	+0.0 ms
CLI --help	86.3 ms	82.8 ms	-3.5 ms

[thymikee]

Reviewed 056b65e. This is a clean, verified pure move of the six adapter leaves into replay-port, with call sites updated and no internal re-exports added. Not blocking: fallow-baselines/health.json still keys the moved session-replay-heal.ts under its old src/daemon/replay/internal path, and AGENTS.md says a rename should move the matching baseline entry, so this can be taken or left. Smoke Tests are still queued rather than failed, and since this diff only renames already-tested modules and touches no runtime dispatch, device command, or daemon route, those lanes exercise nothing this PR changes. No conflicts are known.