Add UI, parser, and session-lifecycle regression tests

[raphant]

Problem

The extension has focused parser tests for Codex and Pi.
It does not have equivalent Claude fixture coverage.
The web interface and the complete session lifecycle have no automated regression coverage.

The current CI workflow runs root tests, extension tests, and type checks.
It does not run the production builds.

Proposed change

Add the smallest test layers that protect runtime parsing, session switching, Review mode, and core keyboard behavior.
Use pure fixture tests first.
Use browser tests only for behavior that crosses components.

Acceptance criteria

• Claude, Codex, and Pi fixtures cover session start, tools, subagents, errors, usage, and completion.
• Session lifecycle tests cover start, update, inactivity, resume, close, and reconnect.
• UI tests cover session switching, panel exclusivity, Review mode seeking, and keyboard shortcuts.
• One browser smoke test verifies the standalone relay-to-dashboard path.
• Continuous integration runs the test and production build commands on pull requests.

Current evidence

• extension/test/pi-session-parser.test.ts covers Pi messages, tools, failures, models, context usage, replay deduplication, and fork-history suppression.
• extension/test/pi-session-watcher.test.ts covers Pi discovery, child sessions, completion, reactivation, and directory precedence.
• extension/test/codex-rollout-parser.test.ts contains focused Codex parser tests.
• web has no test files.
• .github/workflows/ci.yml runs tests and type checks, but it does not run production builds.

Source

• package.json defines the root commands.
• extension/package.json defines the extension checks.
• web/package.json defines the web checks and builds.

[raphant]

Pi regression slice

The first issue #14 slice is on test/pi-relay-regression at 85e5e05.

The new smoke test builds and starts the standalone application in an isolated home directory.
It verifies the configured Pi session directory through the real watcher, relay, HTTP server, and SSE endpoint.
It covers main-session replay, a persisted child agent, an authoritative tool failure, and a later JSONL update without reconnecting.
It uses Node's standard library and adds no dependency.

The pull-request workflow now runs the web, extension, and standalone production builds.
It then runs the Pi standalone smoke against the built package.

All local checks passed:

• Root tests: 36 passed and 0 failed.
• Extension tests: 39 passed and 0 failed.
• Pi standalone smoke: 1 passed and 0 failed.
• Extension source types, extension test types, and web types passed.
• Web, extension, and standalone production builds passed.
• git diff --check passed.

Still outstanding

This slice does not close issue #14.
Claude fixtures, broader lifecycle tests, UI tests, and a browser-level relay-to-dashboard smoke remain outstanding.