b2f2de2bd4d2a0bb9cfefa672b0634f669b4dc7c
7
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Claude
|
4b910f19bf |
🔍 fix: apply the code review — two visible regressions, and the gates behind them
Nine review angles, fourteen verified findings, all but the byte-table generator applied. The two that mattered most were regressions of mine: - a:hover repainted button-styled anchors, and in dark theme --accent-text equals --accent-solid — hovering the hero CTA drew its label in its own background colour. Verified invisible before (computed color == computed background) and distinct after: 10.39:1 dark, 6.90:1 light. The buttons and the skip link now re-assert their label colours on hover. - Footer links had lost every non-hover affordance: the text-decoration:none carve-out plus body-coloured links left a 2.37:1 shade difference as the only cue. The carve-out is deleted — .nav-links a and .brand already declare none themselves — and the accent rule is back on the footer. Also on the page: #ref-filter, the one text input, moves to --border-ui (it still had the 1.68:1 border the token's own comment calls decorative); focusable code surfaces get the --accent-on-code ring at -2px offset, inside the .code overflow clip; #output .out-err drops #ff8095, the last surviving colour of the deleted indigo palette; the two rgba(255,106,126) washes become color-mix over --err-line so a grep for the token finds them. The theme machinery loses a whole block: the dark media query is guarded with :not([data-theme="light"]), so an explicit light toggle falls through to the bare :root palette and the 21-token hand-copy in [data-theme="light"] is gone. Verified in all four system/toggle combinations. The page stops contradicting the repo: it claimed cereale/min "cannot tree-shake — nothing left to shake" while src/treeshake.test.ts proves the opposite on every run. Corrected here and in FRAMEWORKS.md, with the measured figures (1,837 vs 1,996 bytes for one decorator). The release facts the page was hand-bumping — both tgz names, "0.4.0 lives in the repository", the sixty-eight — now fill from meta.js/the bundle like the version badge always has. The workflows close three holes: - The docs sync gate was blind to NEW untracked build outputs (git diff does not report them; demonstrated). Both workflows now run check:docs-sync, one shared script that fails on anything porcelain reports — which also ends the copy-paste divergence between them. - pages.yml deploys on CI succeeding on main (workflow_run) instead of on the push itself, so a deploy implies green tests, not just in-sync docs. The deploy job refuses refs other than main, closing the workflow_dispatch any-branch deploy, and the build job drops pages/id-token — npm postinstall scripts no longer run alongside an OIDC grant. - The Junie action is pinned to the commit behind v1.7.4 rather than the tag, which is the immutability the previous comment promised but a mutable ref cannot deliver. Verified: every fix confirmed in a rendered browser in both themes; 72 contrast pairs still pass; no overflow at 20 widths; 268 tests, build, check:types, check:docs, actionlint all green. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SAcqrz3FcadkYr3xG32CjK |
||
|
Claude
|
5714fab42e |
🧹 refactor: cut the duplicated checks and prose from the tree-shaking work
Acting on ponytail-review. The findings were about verification written twice
and comments restating the CHANGELOG, not about the fixes themselves.
- dist/cjs/package.json had two writers: the build script echoed it, then
build-bundle.mjs rewrote it three lines later with the sideEffects entry.
One writer now, the one that knows what belongs in it.
- Dropped the banner-version assert. Same process, same `pkg.version` going in
and coming out — the "stale bundle" it claimed to catch cannot happen.
- Dropped the `includes('toInstanceSync')` text check. ci.yml imports the flat
bundle and typeof-checks the export, which is the same claim actually tested.
- Dropped the treeshake case asserting annotations survive minification; the
build already asserts it. Its one non-duplicated assertion was the floor on
the expected count, and that gap was real: the build compared flat >= esm, so
if tsc ever stopped emitting annotations both sides would read 0 and the
assert would pass on nothing. Folded in as an explicit `expected < 30` check,
verified by stripping the annotations and watching the build fail.
- Removed the `CEREALE` placeholder from treeshake.test.ts. A template language
for one variable, with two no-op `.replace('CEREALE', 'unused')` calls left
behind by it. Interpolated directly.
- Removed `.replace(/\.js$/, '.js')`, which was the identity function.
- Trimmed the comment above the `Symbol.metadata` install from 16 lines to 7,
and the one above `minifySyntax` from 9 to 3, keeping the parts that are not
written down anywhere else.
Also synced the CHANGELOG's byte table to the README's. The two had already
diverged in the webpack column — which is the duplication the review warned
about, showing up before anyone edited either on purpose.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SAcqrz3FcadkYr3xG32CjK
|
||
|
Claude
|
c48a106a05 |
🌳 fix: make cereale/min tree-shake too, and correct what I overclaimed
An audit of the last commit found two real problems and several claims of mine that went further than the evidence. **cereale/min was not tree-shakable.** scripts/build-bundle.mjs used esbuild's `minify: true`, whose minifyWhitespace pass strips comments — /*#__PURE__*/ annotations included. The published entry point therefore reproduced exactly the bug the previous commit fixed: one decorator came out at 5,066 bytes with all 26 unrelated rule messages, against 1,837 from the per-module entry. Every source-level check stayed green, because they all bundled src/ and the annotations are stripped on the way into dist/. It is now minified for syntax and identifiers but not whitespace: 33.9 KB raw and 9.6 KB gzipped against 26.0/8.7, so about a kilobyte over the wire for a file that behaves correctly however it is used. One decorator via cereale/min is now 1,996 bytes. The build asserts the annotation count survives, and treeshake.test.ts now bundles the published artifact as well as the source — the gap that let this through. **sideEffects was partly inert.** `./dist/cjs/metadata.js` could never match: the build writes dist/cjs/package.json, which becomes the nearest descriptor for everything beneath it, so bundlers read sideEffects from there. That file now carries its own declaration. `./src/metadata.ts` was missing while src/ is published, which declared the Symbol.metadata install droppable in the source tree. Five module-scope caches in utils.ts are annotated for the same reason as the rules. I checked the audit's third blocker — that the Symbol.metadata install is dropped by bundlers — and it is not. It survives every case where it is load-bearing (a decorator import, toPlainSync, modelOf, and a decorated model bundled with an app). It is dropped only when importing nothing but flattenErrors, which needs no metadata, so that is correct. Corrections to my own wording: - "all three bundlers land within 20 bytes" held only for the one-decorator row; larger imports differ by up to a few hundred bytes - "measured through three bundlers, and pinned by a test" read as though the test covered all three; it covers esbuild, on source and on dist - "every rule is a top-level call" — 30 of the 68 are - the docs page "loads nothing from the network" — it fetches its own vendored compiler, same-origin, on first Run. It has no third-party dependencies, which is the claim I should have made - cereale/min's size, everywhere it appears Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SAcqrz3FcadkYr3xG32CjK |
||
|
Claude
|
b4f0657d09 |
📦 feat: add cereale/min, and a verified framework guide
**cereale/min** — the library flattened into one minified ES module, 25.5 KB / 8.6 KB gzipped, for import maps, <script type="module">, Deno and Workers. Built from dist/esm/index.js, so the decorator lowering and the ES2025 target are tsc's; esbuild only flattens and minifies. It is an addition rather than a replacement, and the measurement is the reason. Bundled through esbuild the flat and per-module builds land within 2 bytes of each other; through rollup + terser the flat one is 165 bytes smaller; unused decorators tree-shake out of both. With the size argument a wash, per-module stays the default import for the one thing it does better — readable stack traces without source maps. (My first pass at that measurement reported "shaken" for every symbol because both rollup builds had failed and grep was reading missing files as absence. The check now asserts the bundle is non-empty and that a *used* symbol is present, so it can tell a real result from a broken harness.) **FRAMEWORKS.md** — a recipe per framework, each one run before it was written, with the versions and date verified against. Angular works, which was not obvious: the CLI scaffolds experimentalDecorators: true, but ngtsc erases @Component and @Injectable into static properties rather than leaning on TypeScript's decorator emit. Flip the flag and both systems coexist. Verified with ngc on Angular 21.2 with strictTemplates — templates still type-check and a wrong cereale rule is still TS1240 inside the Angular build. Next.js cannot work inline, structurally: it derives both the SWC parser's decorator support and the transform mode from the one flag, so on gives legacy emit and off makes @ a syntax error. NestJS cannot either — its DI needs design:type from emitDecoratorMetadata. Both have the same answer: keep the cereale classes in a package compiled by tsc and import the built output. Verified inside a program with BOTH legacy flags on, alongside @Injectable() — mapping and validation work, and the compile-time guarantee still holds where the rules are written. Also verified: Bun 1.3 needs no configuration, and a real Vite 8 build with the plugin works where the same build without it silently leaves decorator syntax in the bundle. Version 0.4.0: cereale/min is a new public entry point, and cutting a minor keeps the existing v0.3.0 tag meaningful instead of force-moving it onto a commit it was never cut from. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SAcqrz3FcadkYr3xG32CjK |
||
|
Claude
|
fd42675d3d |
🔗 docs: link the live site, now that Pages is confirmed
Pages serves docs/ from main, so the page rebuilt in #6 is live at avalon-vanguard.github.io/cereale. The README pointed at the local file because the URL could not be verified from here; it now links the site and keeps the local instructions as the fallback. package.json homepage moves there too — npm renders it as the package's headline link, and a live playground is a better landing spot than an anchor inside the README. Adds canonical and Open Graph tags. No og:image: a preview card with a broken image is worse than one without, and there is no artwork yet. check-docs.mjs flagged the canonical link as a remote subresource, which it is not — the browser never fetches it. Rather than exempt the URL, the check now looks at rel and only flags the relations that actually fetch or connect. Verified it still catches a CDN stylesheet, a preconnect and a script src; a check that cannot tell a declaration from a request is one that gets switched off the first time it is wrong. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SAcqrz3FcadkYr3xG32CjK |
||
|
Claude
|
e626a1003a |
🔧 fix: generate .nojekyll instead of hand-editing generated files
The previous commit appended the .nojekyll rationale to docs/vendor/README.md — a file whose own first line reads "Generated by npm run build:docs. Do not edit by hand." CI's "Landing page bundle is in sync with src/" step regenerated it, the text vanished, git diff was non-empty and all three Node jobs failed. The guard did its job; I was the one who put a hand-written paragraph in a generated file. The note now lives in the generator, and build:docs writes docs/.nojekyll itself so it is part of the generated set rather than a loose file that a docs/ rewrite could drop. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SAcqrz3FcadkYr3xG32CjK |
||
|
Claude
|
c828f5cfe9 |
🌾 feat: rebuild the landing page, and stop it from rotting again
The old page had been quietly broken for some time. It loaded
@babel/standalone from an **unpinned** CDN URL, which rolled over to Babel
8 and dropped the `proposal-class-properties` plugin the page asked for, so
Babel.transform threw before it ever reached the decorators — and the
decorator config it passed was `{ legacy: true }`, which 0.2.0 had already
made wrong. Nothing on the page said so. The copy was still selling the
0.1.0 pitch ("Spring-like"), listed about half the decorators, showed
`npm install cereale` for a package the registry returns 404 for, and
claimed "Zero overhead" against a README that publishes the real
microsecond costs.
The rebuild is one self-contained page: hand-written CSS, no Tailwind CDN,
no CodeMirror, and a vendored compiler pinned by package.json. It loads
nothing from the network. The playground runs the real bundled library
across six examples, all verified in a headless browser. The reference
covers all 68 decorators and the full API, counted from the bundle at
runtime so it cannot drift.
The hero's compiler error is not typed into the HTML. scripts/build-docs.mjs
compiles the snippets with the real tsc and writes the verbatim diagnostics
into docs/diagnostics.js, failing the build if a snippet the page calls a
compile error ever compiles — and two snippets that must compile guard
against the harness passing vacuously.
Three guards keep it honest, all wired into CI:
- check:docs fails on any remote subresource
- build:docs + git diff fails if docs/ is stale against src/
- check:types compiles a consumer against dist/ with no DOM lib, no
@types/node and no skipLibCheck
That last one found a real packaging defect: `fromRequest` was declared as
taking the global `Request`, so cereale's own published .d.ts raised
"Cannot find name 'Request'" in any project whose lib and types did not
happen to supply it — inside a dependency, in code they may never call, and
unfixable from the outside. It now takes a structural JsonBody, which a
Request still satisfies. The library's own type tests had been hiding it by
enabling both DOM and skipLibCheck.
An adversarial review of the finished page caught four more: the lede
claimed *every* rule is type-checked (@IsDefined and @IsNotIn deliberately
are not), the guarantee section was wrong about the mechanism (a legacy
decorator does get design:type under emitDecoratorMetadata — the real claim
is about its type signature), one sample called a Movie method on a Media[]
and did not compile, and "nested objects come back as real classes" omitted
that you have to declare them. WCAG contrast was measured rather than
eyeballed: seven real failures fixed in the two themes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SAcqrz3FcadkYr3xG32CjK
|