feat(tui): render think→reply pairs as independent cards

[jkyberneees]

What

Long turns (e.g. "scan this repo") previously merged every response segment into one trailing answer card inside a single turn box, decoupling prose from the reasoning that produced it and producing unexpected interactions on long turns.

Response text is now a first-class timeline kind. Each token/token_delta stream opens — or extends — a reply segment, so every think→reply cycle renders its own raised card in chronological order:

⬡ odek  ⚡2.3s · ⌂12k · ⎇450
  … first thought
  ┌───────────────────────────┐
  │ Starting the scan.        │
  └───────────────────────────┘
  ▶ ✓ 🔍 search_files TODO
  … second thought
  ┌───────────────────────────┐
  │ Found one TODO.           │
  └───────────────────────────┘

Details

• turnItem gains a reply kind with a per-segment glamour cache; resize re-renders segments.
• appendReply keeps msg.content as the \n\n-joined blob (export/stats/hand-built compat); distinct cycles join with a blank line.
• setTurnMarker attaches **Cancelled.** / interrupt / error markers to the last reply segment so they land on the final card.
• closeTurn shared by finalize + session replay; replayTranscript now builds one thinking+reply pair per persisted record — resumed sessions render identically to live ones.
• Turn-body assembly switched to an exact line slice: mouse hit-testing refs land on true header rows (fixes latent misalignment for steps after multiple work items).

Checklist

• make fmt / vet / lint (0 issues) / race tests green
• New regression tests (pairs_test.go): segmentation, marker attachment, residual-blob safety, hit-test accuracy