chore(deps): bump tuika to 0.9 and migrate the dashboard runner

[chaliy]

What changed

serve --tui now runs on tuika 0.9 (from 0.4). The dashboard looks and behaves
exactly as before — same panels, same colors, same q/r bindings — so this is
a no-op for users of the binary. What changed is underneath:

• The host loop moved to tuika 0.9's single FrameSource seam
  (run(theme, async_from_fn(state, view, update))), replacing the four-argument
  run(theme, state, view, update) that 0.9 removed.
• The async update closure now returns UpdateResult (Dirty / Exit / Clean)
  instead of ControlFlow<()>. Ticks and manual refreshes return Dirty so the
  frame still repaints on every poll; unmatched and modified keys return Clean,
  which lets the runner apply its own default input handling.
• Signal handling is split out of the runner closure into an update function.
  That is what makes the tick/refresh/quit contract testable without a terminal —
  previously none of it was covered.
• Import moves: Flex → tuika::components, RatatuiView → tuika::interop;
  RunnerConfig gained a screen_mode field (left at its Alternate default).

ratatui still unifies on 0.30.2 across the RatatuiView interop boundary, so the
widget code is untouched.

Why

tuika 0.4 is five releases behind. 0.9 is a breaking release that collapses the
runner's ten run* methods into one seam and changes the update-result type, so
the bump could not be a version-only change — taking it now keeps the dashboard
on a supported API instead of accumulating more drift.

Before / After

The rendered dashboard is unchanged. Captured live against a running server
(serve --tui) on this branch, after driving traffic through the OpenAI and
Anthropic endpoints:

 ┌────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┐
 │  LLMSim Stats Dashboard   │ ● CONNECTED │ Uptime: 8s                                                               │
 └────────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘
 ┌ Requests ───────────────────────────────────────────────┐┌ Tokens ─────────────────────────────────────────────────┐
 │Total Requests                    6                      ││Prompt Tokens                     57                     │
 │Active Requests                   0                      ││Completion Tokens                 644                    │
 │Completions API                   3                      ││Total Tokens                      701                    │
 │Responses API                     0                      ││Tokens/sec                        0.0                    │
 │Messages API                      3                      ││                                                         │
 │Streaming                         0                      ││                                                         │
 │Requests/sec                      1.20                   ││                                                         │
 └─────────────────────────────────────────────────────────┘└─────────────────────────────────────────────────────────┘
 ┌ Latency ────────────────────────────────────────────────┐┌ Errors ─────────────────────────────────────────────────┐
 │Average                      423.81 ms                   ││Total Errors                      0 (0.0%)               │
 │Minimum                      301.58 ms                   ││Rate Limit (429)                  0                      │
 │Maximum                      566.21 ms                   ││Server (5xx)                      0                      │
 │                                                         ││Timeout (504)                     0                      │
 └─────────────────────────────────────────────────────────┘└─────────────────────────────────────────────────────────┘
 ┌ RPS (current: 1.20, max: 2.00) ────────────────────────────────────┐┌ Models (top 8) ──────────────────────────────┐
 │                      ██  ███  ███                                  ││███ ███                                       │
 │                      ██ ▄███▆▆███▄▄▄▄▄                             ││███ ███                                       │
 │                      ██ ██████████████▄▄▄                          ││███ ███                                       │
 │                   ███████████████████████                          ││███ ███                                       │
 └────────────────────────────────────────────────────────────────────┘│███ ███                                       │
 ┌ Tokens/sec (current: 0, max: 618) ─────────────────────────────────┐│███ ███                                       │
 │  █▄  ▃ ▅▅                                                          ││█3█ █3█                                       │
 │  ██  █ ██                                                          ││cla gpt                                       │
 └────────────────────────────────────────────────────────────────────┘└──────────────────────────────────────────────┘
  q  Quit   r  Refresh

Live-refresh (r) and quit (q) were both exercised interactively: the panels
update out of the tick cadence on r, and q tears the alternate screen down
and exits cleanly.

Before the bump, the crate did not build against tuika 0.9 at all — 6 compile
errors across the two TUI modules (unresolved Flex/RatatuiView, the missing
RunnerConfig::screen_mode field, and the changed run signature).

Risk

• Low
• The blast radius is the optional tui feature; the server, APIs, and library
  paths are untouched. A behavior regression would be confined to the dashboard's
  repaint cadence or key handling, both of which are now covered by tests and were
  verified interactively.

Checklist

• Tests added or updated — 5 new tests over the migrated signal handling
  (tick polls and repaints, dead server records the disconnected state, r
  refreshes out of cadence, q/Esc exit, unmatched and modified keys stay
  Clean), against a stub stats endpoint so no terminal is needed. Full
  suite: 289 passing.
• Backward compatibility considered — no public API or wire-format change;
  dashboard rendering is byte-identical.

---

Generated by Claude Code