feat(ui): turn-scoped thinking indicator with live reasoning readout

[lavindeep]

Branch 5/9 — the flagship of the UI v2 rework. Surfaces what a model turn is doing (the rework was motivated by cloud reasoning models running ~80 min invisible).

What

On submit the app echoes ❯ <message>, then a live frontier line renders as the last transcript row:
✈·· cruising… 38s · 1.8k reasoning — plane glides, phrase advances through flight phases, timer ticks, reasoning-token estimate climbs while the model thinks. Completed tool calls/responses append above it, so it moves down as work finishes (and the auto-following pane keeps it visible). At turn end it freezes to a permanent record: ✓ done · 38s · 1.8k reasoning · 2.4k total.

• Continuous across the whole turn — started once at the first model call, never reset per tool call (unlike v0.10.1's per-call spinner); only turn_finished freezes it.
• Reasoning = chars/4 estimate (no tokenizer); done-line total = exact summed output_tokens from turn stats; elapsed from the runtime's authoritative stats.elapsed_s.
• Gated on [ui] show_reasoning_summary (default on; off → plane/phrase/timer only) — previously dead config, now consumed at both AppUI construction sites.
• Animation via Application(refresh_interval=0.1); the indicator render bypasses the width cache while active. A dangling indicator from an errored turn is discarded at the next turn's start so it never poisons the following timer (full turn-failure UX → branch 6).

Verification

• Gate green: ruff + mypy --strict, 1357 passed.
• Reviewed by two independent reviewers (Opus skeptical + Codex). The continuity guarantee is proven red-first (break begin_response to restart, or end_response to clear → continuity tests red-fail). The cross-turn-poisoning fix (Opus catch) is also red-first tested.
• Default TerminalUI REPL byte-identical; stream_thinking stays a no-op there. DESIGN §31.14 in the same commit.
• Live gate pending (maintainer-run): the animation + reasoning-climb need a real reasoning model — run next.