docs(plans): add escalation-mode wedge design + test plan

Captures the GTM thesis, premises, reduced-scope engineering plan, locked UI
specs, and embedded review report for the Escalation Mode wedge — output of
/office-hours, /plan-eng-review, /plan-design-review, and /codex review.

Codex review surfaced two corrections we applied:
- two-metric framing (manual baseline vs in-product time-to-first-action)
- claim role gate moved in-scope (was deferred TODO)

TODO updates: peer-tech escalation + claim role gate captured (the latter then
moved in-scope by the codex pass).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

.ai/TODO.md

@@ -15,3 +15,9 @@
 - [ ] **Per-test transactional rollback in `test_db` fixture.** Bigger engineering than xdist (which we already shipped). Instead of `DROP SCHEMA public CASCADE` per test, wrap each test in a savepoint and rollback at teardown. ~30-40% additional speedup on top of xdist for test-DB-heavy tests. Real refactor; only worth it if the suite gets significantly larger or runs more frequently.
 - [ ] **Consider `pytest-testmon` for PR-time test selection.** Tracks which tests touched which source files and only re-runs affected ones. Best for small PRs touching ~few files. Adds cache-invalidation complexity; only worth it if the suite stays painfully long even after xdist.
 - [ ] **AssistantChatPage `currentChatRef` guard is a silent return** — `handleSend`, `handleTaskSubmit`, `selectChat`, `refreshFacts`, `refreshActiveFix`, and `refreshPreview` all bail with `if (currentChatRef.current !== sentForChatId) return` when stale. This is by design for chat switching, but it also silently masked the prefill-ref bug fixed in PR #153 — the user just saw "no AI response" with no log, no toast, no Sentry event. Either (a) log a `console.warn`/Sentry breadcrumb on the mismatch path so future drift is visible, or (b) split "expected stale" (chat switch) from "unexpected stale" (ref never updated) so only the latter alerts. Pair with an audit of every `currentChatRef.current = ...` assignment vs every `setActiveChatId(...)` call to make sure they're paired everywhere.
+
+- [ ] **Allow peer-tech to escalate a colleague's session.** Today `POST /ai-sessions/{session_id}/handoff` in [endpoints/session_handoffs.py:48](backend/app/api/endpoints/session_handoffs.py#L48) filters by `AISession.user_id == current_user.id`, so only the session owner can escalate. Real MSP shops have peer hand-offs: Junior A is on lunch, Junior B sees the session is stuck and should be able to escalate it. Auth tweak: switch from session-owner check to `require_engineer_or_admin` + same-account scope. Add a `handed_off_by` audit column (already exists on `SessionHandoff`) so the original-owner-vs-actual-escalator distinction is preserved. Surfaced from /plan-eng-review on the Escalation-Mode wedge plan; v1 wedge demo doesn't need this (solo-founder pilot), but capture for v2 once 3+ pilots are live and a peer-claim need surfaces.
+
+- [ ] **Mobile/responsive design for EscalationQueue + handoff-context screen.** Pre-PMF wedge demo targets desktop only — MSP techs work on laptops/desktops in shop environments. Once 3+ paying customers exist and a tech requests mobile (likely on-call use case), spec the responsive behavior: stacked card layout below `sm:` breakpoint, full-bleed handoff-context overlay on mobile, swipe-to-claim gesture instead of Pick Up button. Surfaced from /plan-design-review on the Escalation-Mode wedge plan.
+
+- [ ] **(MOVED IN-SCOPE for Escalation Mode v1, 2026-04-27)** ~~Add role gate to handoff claim endpoint.~~ Codex review correctly flagged this as wedge-relevant (the race-condition story depends on auth gating). Now part of the Escalation Mode v1 build, not a deferred TODO.