Refresh browser skills and policies; fix navigation test teardown

[ecochran76]

Agents need current runtime/profile selection and resource-ownership guidance while this fork stays on v0.2.9. This refreshes the canonical and installed skills, rebuilds the installed CLI with that same embedded skill, and upgrades the pinned policy-selector from v0.1.22 to v0.1.26. It also repairs the navigation test teardown failures discovered during validation.

Browser guidance covers bounded tab use, owned-resource cleanup on completion/failure/cancellation, OS process verification, and global daemon idle/stop semantics. Policy changes preserve the fork's Graphiti and PR/issue contracts while updating planning/worktree/validation modules and adding provisional tool-first model routing.

The two navigation fixtures waited for HTTP server shutdown before stopping the browser; preconnected/unfinished sockets could block that shutdown. A shared test-only helper stops accepting connections and closes remaining HTTP connections. Socket regressions fail under the old graceful-only behavior and pass after the fix. No production runtime implementation, test skips, or larger timeouts are introduced.

Fixes #5. Plans 0004–0006; decisions and evidence: docs/dev/notes/0002-2026-09-22-policy-v0126-upgrade.md, docs/dev/notes/0003-2026-09-22-navigation-teardown-fix.md.
Base: 6f9e0f1e4a9ce34d99ecde0577c5644d7241a17d. Head: a24fb52b6236e288dedb3837fa4816a447c62eee.

Validation:

• Before fix: local 173 assertions passed, but two navigation teardown hooks timed out at 180 seconds. Instrumentation identified HTTP close as the stalled step; socket-level regressions reproduced it. Original failure receipts remain committed.
• After fix, configured Chromium 150: affected suites 32/32 passed (17.83s); full suite 175/175 across 21 suites passed (26.95s).
• Full run monitor: 27.63s wall, sampled aggregate peak RSS 4.55 GB with two workers; no surviving owned PIDs. Follow-up focused profile census also passed all 32 tests and found zero affected browser-profile processes.
• TypeScript, formatting, Cargo build, and policy audits passed. Prior upgrade validation also passed 12 Rust tests and 22 planning-auditor tests.
• Installed v0.2.9 CLI embeds the exact canonical skill and all three installer targets match. Daemon PID, daemon/sandbox payloads, and browser config stayed unchanged; backup and hash receipts are committed.

No v1 migration or Chromium/default-browser change. Test-only teardown repair needs no new runtime installation. Graphiti feedback mirroring timed out; exact lookup found no episode, so persistence is not claimed.

CI passed on this exact head: Linux/Windows daemon and Rust jobs, bundles, formatting (run 35758275735), and policy (run 35758275680). Optional codesmith was skipped. Approving maintainer review remains the integration gate. No self-approved merge.

[ecochran76]

Reviewed head a24fb52 against base 6f9e0f1. No blocking findings.

The test-only HTTP cleanup stops acceptance before closing existing connections and awaits shutdown completion. Both navigation fixtures use the helper, and real-socket regressions cover preconnect and unfinished-request hangs. Local configured Chromium 150 passed all 175 tests; process evidence shows no remaining test resources and unchanged existing daemon/configuration. Current Linux/Windows CI, formatting, bundles, and policy checks are green. Skill/policy changes preserve v0.2.9 and ownership boundaries.

The operator explicitly instructed “merge if you approve.” Proceeding with a manual merge under that authorization after this review. This records an agent review of its own implementation, not an independent GitHub maintainer approval; no admin bypass or automatic merge is requested. Graphiti feedback persistence remains unverified as documented.