chore(release): v2.7.0 — AI polish history suite

[StevenACZ]

Summary

• Reasoning effort control — global Auto/Off/Low/Medium/High picker in Settings → AI Polish (default Off). Fixes truncated polishes on reasoning models (Mercury 2 spent the whole output token cap thinking); explicit levels reserve extra headroom, rejected params retry once without it.
• Polish with any configured model — the History "Polish with AI" button is now a menu of every configured endpoint's current + recent models; re-polish overrides the provider for that one run.
• Polish history per entry — schema v4 polish_versions table: every applied polish is preserved (model, time, text) and shown as a trail in the History detail. Existing entries backfilled with their latest polish.
• Open in History from the result pill — new button on the completed pill jumps to that dictation, already selected.
• Settings: microphone card switches now align to the card's right edge like every other switch.

Testing

• make ci-check green (lint, secrets scan, Debug build, 255/255 tests — 2 new for the polish trail).
• make release-check green (Release build + size check).
• Schema v4 migration verified against a copy of a real 491-row database (149/149 applied polishes backfilled).
• Reasoning fix verified live against OpenRouter (mercury-2: finish_reason length → stop).