fix: give the published declarations explicit .js extensions

[soimy]

Closes the node16/nodenext item in DEFERRED_WORK.md: the published declarations were written in a form only a tolerant resolver accepts, so the strictest consumers failed inside this package's own files.

The defect, measured

Compiling the documented imports by package name against the built package, before the fix:

consumer mode (skipLibCheck: false)	before	after
bundler (Vite/webpack)	pass	pass (gated)
node16	fail, 5 errors	pass (gated)
nodenext	fail, 5 errors	pass (gated)
nodenext with skipLibCheck: true	pass	pass
node10 + CommonJS	pass	pass (measured by hand)
node10 + CommonJS on TypeScript 4.6	pass	pass (measured by hand)

dist/index.d.ts(1,39): error TS2834: Relative import paths need explicit file extensions in
  ECMAScript imports when '--moduleResolution' is 'node16' or 'nodenext'.
dist/index.d.ts(2,56): error TS2835: … Did you mean './maxrects-packer.mjs'?

The cause is structural: @rollup/plugin-typescript emits one declaration per source module and keeps the relative specifiers extensionless, while the package is "type": "module".

The fix

scripts/fix-declaration-extensions.mjs, run from postbuild, rewrites relative specifiers in dist/**/*.d.ts to end in .js — the ESM-correct form: TypeScript maps ./x.js to ./x.d.ts, and the runtime artifacts are self-contained bundles with no relative imports, so the extension only ever has to resolve as a declaration. 17 specifiers in 5 files; no .js/.mjs/.cjs artifact is touched.

The script is defensive in the two ways that matter for a build step: it refuses to write a specifier that has no sibling declaration to resolve to (that would ship a broken path silently), and after rewriting it asserts nothing was left extensionless. Both were exercised on purpose — export * from "./missing" produces Error: …/dist/__probe.d.ts: "./missing" has no sibling declaration to resolve to. A second run is a no-op.

The gate

scripts/verify-package.mjs now compiles the type fixture under bundler, node16 and nodenext, with skipLibCheck: false, in a consumer project marked "type": "module" so the two strict modes read the fixture as ESM — so every mode this description claims is a continuous CI gate, not a one-off measurement. Teeth verified per entry by removing one .js from dist/index.d.ts: the gate prints the TS2834 message and exits non-zero, and because node16 is the first strict mode it reaches, the failure is reported as under node16. node10 stays ungated on purpose — it never needed the extension, and a wrong .mjs fix would only show up there (see AGENTS.md).

Compatibility

The extension is not a node10 problem: that resolver maps ./x.js to ./x.d.ts too, which the table above pins with TypeScript 4.6 — the oldest version anyone is realistically still building against. No exports field is involved, so deep imports keep working exactly as before.

AGENTS.md documents the step, the matrix and the "three gates" wording; the item is removed from DEFERRED_WORK.md. Local gates: lint 0/0, format:check, typecheck, npm test (98 passed / 2 skipped, exercising the postbuild step end to end), cover — still 100% on statements, branches, functions and lines — and verify:package.

[greptile-apps]

Confidence Score: 5/5

[High risk] Build process now rewrites declaration files after compilation.

No outstanding findings block merging.

Summary

The PR rewrites relative imports in published declarations to use .js extensions and verifies the packed package’s types under bundler, node16, and nodenext. No outstanding findings remain.

_{Reviews (2) · Last reviewed commit: "test: gate node16 in the package type fi..."}