feat: add iterative refinement to Reply Drafter (#2298)

[OwenMcGirr]

Summary

Reply Drafter was one-shot today — open it, get 3-5 suggestions, tap one. When none fit, the only escape was to close and reopen and get a fresh roll of the dice. This adds a text field + Regenerate button at the bottom of the suggestions list for iterative refinement.

The key architectural choice: refinement sends the model the image + the previous reply set + the user's note, framed as "Here's what you wrote, here's what the user wants different, produce a new set." The note can't be mistaken for the message to reply to because the actual conversation is in the image and the previous replies clearly bracket what's being refined.

• First open is unchanged. Empty previousReplies collapses the prompt to the original RULES_BLOCK + FORMAT_BLOCK, byte-for-byte identical to today.
• Subsequent Regenerates carry the latest set forward, so you can iteratively tighten ("be more formal" → "now shorter" → "drop the question one").
• Failed-state Retry uses fresh draft(), not refinement — nothing useful to refine against if generation failed.
• Bottom placement is deliberate for switch-access: suggestion cards stay the primary scan targets; the field is the last scan stop and can be skipped entirely.

Files:

• service/llm/ReplyDrafterTask.kt — object → data class with previousReplies + refinement. Computed prompt inserts a REFINEMENT block between the rules and format blocks when previousReplies is non-empty. Format block stays last (recency benefit preserved).
• screens/replydrafter/ReplyDrafterViewModel.kt — tracks lastReplies, exposes setGuidance(text) (200-char cap), regenerate(). draft() stays fresh; regenerate() refines.
• screens/replydrafter/ReplyDrafterScreen.kt — pinned GuidanceInputRow (OutlinedTextField + Refresh IconButton), visible in Suggestions and Failed states.
• screens/replydrafter/ReplyDrafterActivity.kt — observes guidance state, wires setGuidance and regenerate callbacks.
• 2 new strings; 14 unit tests pass (11 parser unchanged + 3 new prompt-shape).

Closes #2298

Test plan

• ./gradlew compileDebugKotlin testDebugUnitTest installDebug
• Manual on S26 (AICore): first draft unchanged; "be more formal" produces formal alternatives; clearing the field still re-rolls
• Manual: iterative chain ("be more formal" → "now shorter" → "drop the decline one") visibly refines step by step
• Manual: MediaPipe fallback on a non-AICore device behaves the same (with weaker instruction-following — expected)