Follow-up to the ES5 syntax fix, closing the gaps that investigation left.
An ES5-only runtime sandbox. The acorn check catches syntax, but syntax is
only half the contract: preset-env lowers syntax and never polyfills library
calls, so one `Array.prototype.includes` in the parser source compiles
cleanly, passes every test on modern Node, and throws on the browsers es5.js
exists for. The new test runs both legacy bundles in a vm context with the
post-ES5.1 globals, statics and prototype methods deleted, and asserts
bundled.js additionally installs the polyfills its README entry promises.
Includes a test that the sandbox really strips, so it cannot quietly pass
against a modern global.
A consumer type-check across every module resolution mode, run in CI against
the packed tarball. attw already checks that types *resolve* per condition;
it compiles nothing, so it cannot catch a declaration that resolves correctly
and then misdescribes the runtime. Negative cases are asserted too — the maps
must stay non-importable as named exports, which is the line index.d.mts
draws deliberately and only a failing compile can hold.
Also documents two findings that were investigated and deliberately left
alone: the bundled.js size increase is the core-js 2 -> 3 upgrade rather than
waste, and `useBuiltIns: 'usage'` would shrink it by breaking the documented
"all needed polyfills" contract; and the src/*.js ESM-in-CJS wart (publint
warnings, Yarn PnP, Node < 20.19) is longstanding and identical on 2.14.1,
with the nested-package.json fix blocked on @babel/register.
Verified by breaking each guard in turn: an ES6 API call injected into es5.js
fails the sandbox test, and an index.d.mts with its `parse` export removed
fails all three ESM resolution modes while the CJS modes correctly still pass.
`bundled.js` stopped being ES5 when the webpack build was replaced by tsdown
in #628. It parses at ecmaVersion 2015 but not 5:
var t=(t,e)=>()=>(e||(t((e={exports:{}}).exports,e),t=null),e.exports)
That is rolldown's `__commonJS` interop helper. `@rolldown/plugin-babel` only
transforms input modules, and rolldown appends the helper afterwards; terser
with `ecma: 5` avoids introducing newer syntax but does not transpile, so the
arrow functions reached the published file. `es5.js` has no CommonJS
dependencies and never gets the helper, which is why only `bundled.js` broke.
The effect is total rather than partial: in an ES5 engine the whole script is
a SyntaxError, so `bundled.js` — the bundle that exists specifically to serve
those engines, polyfills included — does not load at all there.
Lower the emitted chunk with a babel renderChunk pass that runs after bundling
and before terser, so rolldown's own helpers are covered too. Costs 2.8 kB
(+1.6%) on bundled.js; es5.js is unchanged at 34 kB.
The existing guard was a grep for backticks, which this syntax slips straight
past. Replace it with an acorn parse at ecmaVersion 5 over both legacy
bundles, plus a tokeniser check for real template literals (backticks inside
core-js string literals are fine, and 2.14.1 shipped three of them).
Verified by reverting the build fix: `bundled.js parses as ES5` fails, and
passes again once restored. Also confirmed bundled.js loads and parses the
live navigator.userAgent in a real browser.
Two follow-ups to the dual packaging change (#628), found while running a
consumer-facing regression sweep against published 2.14.1.
An exports map is a closed list. Before #628 bowser had no exports map, so
every published file was reachable by subpath; afterwards LICENSE, README.md
and index.d.ts resolved to ERR_PACKAGE_PATH_NOT_EXPORTED. Nothing in the
documented API regressed, but `/// <reference types="bowser/index.d.ts" />`
and tooling that resolves the license by specifier both did. Add the four
non-code files to the map so the published surface matches 2.14.1 exactly,
and assert it in the package smoke test (25 -> 29 checks).
The acceptance suite asserted src/bowser.js and es5.js against the UA corpus
but not bundled.js or bowser.mjs. bowser.mjs is what the `import` condition
resolves to, so it is the file every modern ESM and bundler consumer runs,
and a build regression confined to it would have kept CI green. Assert all
four artifacts against the spec instead.
Verified by injecting a Chrome-parsing regression into bowser.mjs and into
bundled.js separately (each turns the suite red, labelled by artifact), and
by running the new smoke assertions against a pre-fix tarball (exit 1) and
this one (exit 0) on Node 12.16, 14, 18, 20 and 22.
