chore: sync upstream production and isolate fork regressions

[henry701]

[ai-agent-generated] This description is LLM-generated.

Summary

• Merge upstream production through f54ce31, preserving the fork's durable steering/queue, Stop, rollback replacement, model selection, and token-accounting integrations.
• Resolve the lockfile conflict and move retained historical-layout regressions into a fork-owned spec instead of conflicting with upstream's retired spec. The shared projection test now follows upstream's new-layout coverage.
• Pin historical-layout tests before the upstream sunset, including the token-usage matrix that otherwise silently tests the new layout twice. No runtime refactor or audit document added.

Validation (Linux)

• All 30 workspace typecheck tasks; app e2e typecheck; generated-client check.
• All 10 configured workspace test tasks, including 3,642 opencode and 1,143 core tests.
• Separate TUI suite: 221 passed, one skipped; real terminal queue-edit/API smoke: three passed.
• HttpApi coverage, authentication, and Effect gates: 236/236 each.
• Full Chromium + Firefox run: 251 passed, three existing skips. Initial startup/hover failures passed in isolated reruns and the final complete run; historical-layout failures were fixed with explicit clocks.
• Frozen lockfile validation with repository-pinned Bun 1.3.14.

Keep merge commits (do not squash) to retain upstream ancestry and avoid re-resolving these conflicts. Windows is non-blocking per the fork owner's direction.

CI outcome

GitHub typecheck, Linux unit/generated-client/HttpApi, and Linux browser jobs passed. Windows unit setup failed before tests: Bun 1.3.14 hit an ENOTEMPTY cache rename and ENOENT while applying the upstream OpenAI 3.0.88 patch. This is a Windows installation failure, not quota exhaustion; it does not block this Linux-qualified merge.