feat: Vite support, template coverage, and ember-cli decoupling - #432

Explorations for #431 — continues the v2 addon migration and takes it to a working state across classic, Embroider, and Vite builds.

What #431 asked for

The issue is really about decoupling from ember-cli, not just "add Vite". Status against the specific asks in that thread:

• Middleware usable directly by testem, not via the ember-cli hook — done. coverageMiddleware and viteTestemMiddleware are plain exports from ember-cli-code-coverage/testem, wired into testem.cjs with no ember-cli involvement.
• Commands working without ember-cli — done. New coverage-merge bin runs the merge standalone, with optional --root/--config.
• Composable subpath entry points (gossi's /istanbul, /vite, /testem, /babel) — done, as /core, /vite, /testem, /babel, plus /glimmer and /merge. Named /core to match the internal directory, consistent with every other subpath here.
• Native V8 coverage instead of Istanbul — not attempted. It's raised as an open question in the thread and nobody committed to it; still Istanbul-based.
• vitest browser mode / ember-vitest — untouched, still experimental upstream.

Regressions fixed from the original modernize commit

These were broken on the branch before this work, not by it:

• Embroider reported no coverage at all. The default exclude changed from */node_modules/**/* to **/node_modules/**/*, and Embroider ≥3.1 rewrites the app into node_modules/.embroider/rewritten-app/, so the app's own files were all excluded. The committed snapshot proves it — it contained one entry, this addon's own file, and none of the app's.
• ember serve and ember test --server lost their fresh-map-per-request behavior, and a project-defined coverageEnvVar was ignored in favour of a hardcoded COVERAGE.
• In-repo addon coverage was missing its addon/ and addon-test-support/ trees.
• This addon's own source leaked into every fixture's coverage report, and the HTML reporter wrote files outside its output directory when those paths crossed workspace boundaries.

Vite

Incorporates the intent of #434, with fixes beyond it:

• Instrumentation moved into a post-enforce Vite transform, so .gjs/.gts are seen as plain JS regardless of plugin ordering.
• Build-time zero baseline, so files no test imports still report at 0% rather than vanishing.
• viteTestemMiddleware for ember test --path dist, where the Vite dev server isn't running.

Template coverage

Related to #433 but a different implementation, because that PR's central assumption no longer holds: it relies on camelCase helper names resolving through the classic resolver, and current ember-resolver removed that fallback deliberately — the runtime warning says so outright. Loose mode needs dash-case, strict mode needs camelCase, so they can't share one set.

• .hbs (loose mode) works with no setup beyond templateCoverage: true in config/coverage.js.
• .gjs/.gts (strict mode, including apps on ember-strict-application-resolver) works on Vite, via createTemplateCoveragePlugin({ strict: true }) plus templateCoverageImportPlugin, which supplies the JS bindings strict mode requires. On classic/Embroider those files get no template coverage — see below.
• Covers {{#if}}/{{#unless}}, {{else if}} chains, {{#each}}/{{#each-in}} with {{else}}, and inline conditionals including attribute position.
• Survives throwUnlessParallelizable, which feature: template coverage #433 doesn't address.

Also fixes adjustCoverage overwriting rather than merging when two raw keys resolve to the same file — the template and JS halves of a .gjs file depend on that merge, and it was already wrong for any other collision.

Tooling

• Dependencies current across root, addon, and every fixture. TypeScript held at 6.0.3 (typescript-eslint caps below 6.1) and Babel at 7.x (the Ember toolchain isn't Babel 8 ready).
• One pnpm lint at the root chains formatting plus every workspace package's own lint, fixtures included (pnpm -r --if-present run lint), in about 6 seconds. Five fixtures had lint failures blocking this — missing type="button", a prettier mismatch, a verbatimModuleSyntax type-only import, and a node/no-missing-require false positive where eslint-plugin-node predates exports maps and can't resolve ember-cli-code-coverage/testem even though node resolves it fine. All fixed.
• CI: action versions bumped, least-privilege permissions, job timeouts, workflow_dispatch, Node 26 added to the matrix.
• Integration tests serialized — run in parallel, real Ember builds starved each other for CPU and coverage collection degraded under load, making snapshots flap.
• Consolidated the two diverged CHANGELOG.md files. The nested one uniquely held v2.0.0-beta.1, v2.0.0-beta.2 and v1.0.0-beta.9 (root skipped straight from beta.8 to v1.0.0); those are merged into root at their chronological positions and the nested file removed. Nothing was dropped from root — every version in the nested file is verified present.

Testing

58 tests across 18 files. Restores the in-repo-addon, custom-path, in-repo-engine, v1-addon and scoped-v1-addon fixtures deleted during modernization (they cover namespace mapping and modifyAssetLocation, the most breakage-prone code here), and adds fixtures for Vite and for template coverage.

Not verified / open

• Strict-mode template coverage does not work on classic/Embroider, and this is now tested rather than assumed. ember-template-imports converts .gjs through its own preprocessor, and the AST plugin registered via setupPreprocessorRegistry never reaches the result; there's no transforms hook on that path to register against, which is why the Vite setup needs hand-wiring in babel.config.mjs and a classic app has no equivalent file for it. Those templates are silently skipped rather than breaking the build — confirmed by compiling a real {{#if}} in a .gjs file. Their JavaScript is still covered.