refactor: consolidate public SDK entry barrels into src/sdk/ — Phase 5

[thymikee]

What

Consolidates the package's public entry points into a dedicated src/sdk/ folder so the public surface lives in one clearly-owned place, per plans/perfect-shape.md §5.5 ("sdk/ = re-export barrels only") and the §3 target DAG. This is the next Phase-5 layering item after remote/ (#951) and client/ (#960), and follows the same behaviorless path-codemod pattern.

The published npm surface is byte-identical — this is pure internal reorganization behind the same public contract.

Barrel structure

All 11 public subpaths (., ./io, ./artifacts, ./metro, ./batch, ./remote-config, ./install-source, ./android-adb, ./contracts, ./selectors, ./finders) now resolve through src/sdk/:

• 9 already-thin barrels moved verbatim (git rename; internal import paths re-depthed by one ../): index, artifacts, metro, batch, remote-config, install-source, android-adb, contracts, selectors.
• io and finders keep their implementation at src/ root (they contain real logic + have internal importers). src/sdk/io.ts and src/sdk/finders.ts re-export them (export *), so src/sdk/ stays barrels-only.

package.json exports diff

None — unchanged. Because each rslib entry key is preserved (index, io, …), dist output stays dist/src/<name>.{js,d.ts}. The published subpaths and files do not move, so no exports/main/types rewrite was needed (same technique #960 used for internal/companion-tunnel). This is the safest possible outcome: the exports map is provably identical.

rslib.config.ts:  index: 'src/index.ts'  ->  index: 'src/sdk/index.ts'   (key 'index' unchanged)
                  io:    'src/io.ts'      ->  io:    'src/sdk/io.ts'      (key 'io' unchanged)
                  ... all 11 public entries; 5 internal/* entries untouched

npm pack --dry-run before / after (proving identical public surface)

	baseline (origin/main)	after
packed files	146	146
package.json exports/main/types	(unchanged)	identical
public .d.ts (all 11)	—	byte-identical
public .js (11)	—	10 byte-identical; selectors.js differs only by 1 internal chunk ref

The only delta in the entire tarball: one internal code-split chunk was renumbered dist/src/9836.js -> dist/src/893.js (byte-identical content; a deterministic rspack chunk-id shift from the module-graph move). Every chunk that imports it just updates that one ./9836.js -> ./893.js reference. No public subpath, symbol, or .d.ts changed.

--- baseline/pack-files.txt
+++ after/pack-files.txt
> dist/src/893.js
< dist/src/9836.js       (same content, renamed)

Repoints (all internal / test — no public change)

• rslib.config.ts: 11 entry sources -> src/sdk/*.ts (keys unchanged).
• 7 public *-test imports -> src/sdk/*.
• 3 internal src importers of the selectors barrel (daemon/handlers/find.ts, commands/interaction/runtime/{selector-read,resolution}.ts) split to import impl directly (utils/selectors-parse.ts + daemon/selectors.ts) — removes internal dependence on the public barrel.
• .fallowrc.json entrypoints + vitest.config.ts coverage exclude (src/sdk/** glob).

Package-subpath consumers (agent-device/io, website client-api.md, installed-package-metro.test.ts) reference the unchanged subpaths — untouched.

Verification

• tsc -p tsconfig.json --noEmit -> 0
• rslib build -> 0; dist entry list identical (98 files), all 22 public entry files present
• npm pack --dry-run -> file list + exports identical (see above)
• oxfmt + oxlint --deny-warnings -> clean
• fallow audit --base origin/main -> no issues in 31 changed files
• vitest run --project unit -> 2883 passed
• (scripts/layering/check.ts not present in repo)

[github-actions]

Size Report

Metric	Base	Current	Diff
JS raw	1.4 MB	1.4 MB	-8 B
JS gzip	450.5 kB	450.5 kB	-11 B
npm tarball	549.9 kB	549.9 kB	-16 B
npm unpacked	1.9 MB	1.9 MB	-8 B

Startup median (7 runs, lower is better):

Scenario	Base	Current	Diff
CLI --version	26.4 ms	27.0 ms	+0.6 ms
CLI --help	47.1 ms	47.8 ms	+0.7 ms

Top changed chunks:

Chunk	Raw diff	Gzip diff
dist/src/cli.js	-1 B	-3 B
dist/src/9722.js	-1 B	-2 B
dist/src/selector-runtime.js	-1 B	-2 B
dist/src/session.js	-1 B	-1 B
dist/src/2577.js	-1 B	-1 B

[thymikee]

Reviewed #986 against plans/perfect-shape.md Phase 5 §5.5 and the package-surface constraints. I don't see actionable blockers: this is a public SDK barrel relocation, the rslib entry keys remain unchanged, package.json exports are untouched, src/sdk/** stays re-export-only, implementation-bearing io and finders remain behind SDK re-export barrels, and internal command/daemon code no longer depends on the public selectors barrel. Checks are green, including Typecheck, Unit Tests, Integration Tests, Fallow Code Quality, Layering Guard, Lint & Format, Smoke Tests, Bundle Size, and CodeQL. No live device evidence requested because this PR changes packaging/module locations, not device-facing behavior.

[github-actions]

PR Preview Action v1.8.1
Preview removed because the pull request was closed.
2026-07-01 08:04 UTC