feat(agent): browser-extension verification (load, drive, compile, re-run)

[tcballard]

What & why

Proofkeeper couldn't test a browser extension: it drove a vanilla, non-persistent Chromium page, had no config for an extension, and the emitter always produced a test({ page }) spec on the default fixture — none of which loads an extension. Chromium loads an unpacked extension only in a persistent context with --load-extension, and an MV3 extension's ID is regenerated on every load.

This adds first-class extension verification through the same drive→compile→fidelity→run→write-back loop: load the unpacked extension, let the model exercise its popup/options pages (chrome-extension://<id>/...) and its effect on host pages, and emit a compiled spec that re-loads the extension and re-discovers the ID at run time — so the committed test and the fidelity gate verify the real extension, never a stale recorded ID.

How

• Launch (src/cli.ts browserDrive): when an extension is configured, launch a persistent context with channel: "chromium" and --load-extension / --disable-extensions-except. channel: "chromium" selects Chromium's new headless — the only headless mode that loads extensions (old headless never does; pure headed needs a display). Verified empirically against this repo's pre-installed Chromium.
• ID discovery (src/agent/extension.ts): loadExtension() reads the runtime ID from the MV3 service worker (MV2 background-page fallback, then waits for the worker). The driver surfaces the ID and the chrome-extension://<id>/ base to the model.
• Emitter, the moat (src/compiler/emit.ts): a new mode when RecordedSession.extensionPath is set — emit a persistent-context spec that relaunches with the extension, rediscovers extId, and rewrites recorded chrome-extension gotos to use the runtime extId. Non-extension specs are byte-identical to before (regression-guarded).
• Config/CLI: a --extension <dir> flag on qa, and a per-environment extensionPath, threaded through resolveTarget → QaOptions → DriveOptions for both the qa and scoped paths.

Usage

proofkeeper qa --corpus path/to/rac/ --url http://localhost:3000/ --extension ./my-extension

Or per environment in the config: "environments": { "dev": { "url": "...", "extensionPath": "./my-extension" } }.

Tests & gates

• Unit: extensionIdFromUrl / loadExtension (fake context); emitter extension mode (persistent context, channel: "chromium", runtime-ID rewrite, determinism, non-extension unchanged); config parse + resolveTarget.
• Browser-gated integration (tests/extension.integration.test.ts, PROOFKEEPER_E2E): loads a real MV3 fixture extension (tests/fixtures/extension/), discovers the ID, drives the popup — passing against the pre-installed Chromium. Added to the CI e2e job.
• Full suite 234 passed; typecheck + build clean.
• Dogfood corpus: req-extension-verification + design-extension-verification linked to the autonomous-qa-enhancements roadmap, ## Verified By the new tests. rac validate 23/23, rac relationships --validate 0 issues, rac review 100/100.

Scope

MV3 Chromium extensions loaded unpacked. Packed .crx, the Chrome Web Store, non-Chromium browsers, and a manifest-key stable-ID mode are out of scope (noted in the design). Releasable → next CalVer. Independent of the OpenAI-adapter PR.