fix(app): sync upstream and harden development web session state

[henry701]

[ai-agent-generated] This PR description was LLM-generated.

Scope

Integrate upstream through 70b4ca8c181e4c1ac6d8993b86249d824487ec65 while preserving fork-specific web behavior. Save the session-state, native projection, early Stop, rollback, provider-discovery and CI/test improvements developed during the audit. Use focused adapters and point integrations rather than reverting fork features.

Owner-approved merge with follow-ups

The repository owner explicitly requested committing/pushing all current improvements and merging PRs #7 and #8 now, with remaining issues fixed in follow-up PRs. This supersedes the earlier draft hold. It is not a claim that browser tests or hosted CI are all green.

The owner confirmed that CPU throttling caused the slower tab-switch benchmark timings. The throttled measurements are not comparable with the earlier baseline and do not establish a code-induced performance regression.

Verification

• App: 800 unit tests and 51 browser-environment tests passed; app and enumerated e2e typechecks passed.
• Affected core command/prompt/replacement tests: 38 passed, 100 assertions.
• Full pinned-Bun Linux unit run before final focused additions: 10/10 uncached Turbo tasks passed; core 1,142 tests; opencode 3,627 passed / 22 skipped / 1 todo.
• Seven timeline specs: 43 passed, 3 existing Firefox CDP skips.
• Stop/pending/rollback/transport focused browser run: 24 passed, 2 command Undo/Redo failures.
• Recorded/rollback/live context usage: 6 passed across Chromium and Firefox.
• Fresh request/todo/child/review/onboarding/smoke run: 18 passed, 8 failed.
• Production benchmark: 2 structural cases passed; timing comparison affected by owner-applied CPU throttling.
• Workflow actionlint passed. Normal pre-push typecheck hooks remain enabled.

These are separate, overlapping runs, not an aggregate all-green browser suite. Prior CI failures included real subprocess/portability failures and installation/cancellation problems; quota exhaustion was not established. Windows clean-install validation remains incomplete.

Recorded follow-ups

See docs/fork/web-audit.md for evidence, reproduction commands and exit criteria:

• WEB-01: connection-dialog provider hydration crash after discovering OpenCode Go.
• WEB-02: command palette does not reopen for Redo after Undo in the clean regression.
• WEB-03: review/terminal tree geometry readiness.
• WEB-04: cached first-paint probe counts a replaced text span; distinguish leaf updates from structural remounts.
• WEB-05: full-history smoke context/global-stream console errors.
• CI-01/02: Windows patched Bun ENOTEMPTY, clean Windows tests and hosted CI parity.
• SYNC-01: newer upstream e2894562f8ba943d72172d10b727c24d5f650c16 (console usage normalization/tier configuration) and later changes remain follow-up work.
• RELEASE-01/PERF-01: deployed Basic-auth parity, full exact-head browser runs and matched-CPU performance measurements.

Failing tests were retained, not skipped or weakened. No deployment, service restart or binary reinstall is included.

Post-merge verification — 2026-09-05

Both PRs are merged: production 327ce4d046b14334f9d221c90841bbf7d1c72308, development 1c504d299ca8ded867ee7ebc8a5ca216d83b82d3. Both original pushed improvement heads are verified ancestors of their fork target branches; the resulting dev and production trees are identical.

SYNC-01 update: the fork's upstream-sync workflow merged e2894562f8ba943d72172d10b727c24d5f650c16 into both target branches before these PR merges. Thus the audit file's pre-merge SYNC-01 checkpoint is now resolved for that commit. The merged console changes passed five provider-usage tests, two tier-threshold tests, and both affected console package typechecks. Future upstream updates remain ordinary follow-up work.

Both pre-push hooks passed 30/30 typecheck tasks. Development independently passed the 800 + 51 app tests and e2e typechecking. The final affected core run including MCP passed 39 tests / 125 assertions. Hosted checks were pending at the merge; no hosted all-green claim is made. Remaining WEB/CI/deployment follow-ups are unchanged.