refactor(components): extract DraftResponsePanel + pin DraftTabHandle (Wave 5.7 Phase B+D)

[saagpatel]

What

Completes Wave 5.7 with a narrow Phase B + Phase D pass:

Phase B — DraftResponsePanel extraction. The response + alternatives + saved-suggestion JSX fragment (38 lines in the shell) becomes a dedicated DraftResponsePanel component. Shell reduces from 2138 → 2127 LOC on this PR.

Phase D — DraftTabHandle shape pin. New DraftTab.handle.test.tsx adds compile-time + runtime guards around the DraftTabHandle interface. Eight external consumers import this handle — a type-level drift-detector is the highest-value first test the shell can have.

Why (narrow scope for Phase B)

The Wave 5.7 plan sketched three panels: DraftIntakePanel, DraftResponsePanel, DraftChecklistPanel. In the actual shell:

• Intake already lives inside TicketWorkspaceRail (a separate feature module). Wrapping it in a thin `DraftIntakePanel` would add indirection without reducing shell complexity.
• Checklist JSX lives inside DiagnosisPanel. Same indirection cost.
• Response + alternatives was the only fragment with a real extraction win — 3 sub-components glued together with conditional-render logic.

So this PR extracts the one that earns its keep and documents the decision.

Phase C note

With Phase A (5 hooks) + this PR landed, `DraftTab.tsx` is at 2127 LOC. Further reduction to the plan's 700-LOC target would need a separate sweep of the remaining ~30 handlers (draft save/load, handoff pack copy, similar-case openers, runbook advancers, template modal glue). None of them form a cohesive "one more hook" concern — they're each self-contained and already use extracted hooks/helpers. A 700-LOC shell would require significant JSX restructuring beyond the decomposition pattern used so far.

Testing

• `pnpm tsc --noEmit` — clean.
• `pnpm test` — 205 pass (was 200; +4 DraftResponsePanel tests + 1 DraftTabHandle shape test).
• `pnpm ui:gate:static` — clean (pre-existing warnings only).

Risk / Notes

• `DraftTabHandle` interface unchanged (now regression-guarded).
• Default export unchanged.
• `ResponsePanel`, `AlternativePanel`, `SavedResponsesSuggestion` consumer contracts unchanged — only their call-site moved from `DraftTab` to `DraftResponsePanel`.
• Conditional-render rules (suggestion hidden when response present, alternatives hidden while generating/streaming) preserved exactly.

Wave 5.7 cumulative

• DraftTab.tsx: 2573 → 2127 LOC (-446, ~17%)
• Tests: 184 → 205 (+21 direct DraftTab concern tests; was zero)
• Files added: 5 hooks + 1 panel + 6 test files