refactor(components): extract useDraftGeneration hook (Wave 5.7 Phase A complete)

[saagpatel]

What

Extracts the final Phase A hook, `useDraftGeneration`, from `src/components/Draft/DraftTab.tsx`. Completes the Wave 5.7 hook decomposition:

• `useDraftApproval` ✓ (PR refactor(components): extract 4 DraftTab hooks (Wave 5.7 Phase A part 1) #54)
• `useDraftFirstResponse` ✓ (PR refactor(components): extract 4 DraftTab hooks (Wave 5.7 Phase A part 1) #54)
• `useDraftChecklist` ✓ (PR refactor(components): extract 4 DraftTab hooks (Wave 5.7 Phase A part 1) #54)
• `useDraftIntake` ✓ (PR refactor(components): extract 4 DraftTab hooks (Wave 5.7 Phase A part 1) #54)
• `useDraftGeneration` ✓ (this PR)

File size: 2283 → 2138 LOC (-145 on this PR; -435 total across both PRs from original 2573, ~17% reduction).

Why

`handleGenerate` + `handleGenerateAlternative` + `handleChooseAlternative` + `handleUseAlternative` form one cohesive concern: streaming LLM response generation and the alternatives lifecycle. The hook owns the two loading flags (`generating`, `generatingAlternative`) and the `firstDraftStartMsRef` time-to-draft tracker, accepting setters for the response-content state (`response`, `sources`, `metrics`, `confidence`, `grounding`, `originalResponse`, `isResponseEdited`) that stays in the shell because many non-generation handlers also write those fields.

How

1. Create `useDraftGeneration.ts` with the 4 callbacks + 2 flags + the first-draft-timestamp ref.
2. Add `useDraftGeneration.test.ts` (4 `renderHook` tests covering: no-model block, success path, alternative skip, alternative use).
3. Update `DraftTab.tsx` to consume the hook, removing 4 inline `useCallback` handlers and the two `useState` + `useRef` declarations they owned.
4. Replace the `setGeneratingAlternative(false)` + `firstDraftStartMsRef.current = null` lines in `handleClear` with a single `resetGeneration()` call.

Deferred to a follow-up PR

• Phase B — sub-component panel extractions (`DraftIntakePanel`, `DraftResponsePanel`, `DraftChecklistPanel`).
• Phase C — shell reduction to ≤ 700 LOC (depends on Phase B).
• Phase D — DraftTab integration tests (depends on Phase C).

Testing

• `pnpm tsc --noEmit` — clean.
• `pnpm test` — 200 tests pass (was 196; +4 new hook tests).
• `pnpm ui:gate:static` — clean (only pre-existing warnings).

Risk / Notes

• `DraftTabHandle` interface unchanged.
• `useWorkspaceDraftState` consumer contract preserved (shell still owns the response-content state and passes its setters in).
• The conversation-submit handler in the shell still calls `setGenerating(true/false)` directly — exposed from the hook for that exact purpose.