TUI: Gemini response mode breakdown and drift detection

[coreyt]

Problem

The state store tracks gemini_outcomes (text, thought_only, tool) per client and per provider. The Overview screen shows raw counters in detail panes, but there's no way to:

• See the distribution of response modes over time
• Detect drift in thinking-mode behavior (e.g. a model suddenly returning thought_only for prompts that used to produce text)
• Correlate Gemini response mode with error rates

Scope

• Mode distribution in provider and client detail panes — show as percentages rather than raw counts (e.g. "text: 72%, thought_only: 23%, tool: 5%")
• Drift indicator — if the ratio shifts significantly from the rolling average, surface a warning (could be an alert engine rule)
• Correlation view — in the Guards screen detail, show Gemini response mode alongside composite score to help operators spot patterns

Backend references

• fast/state.py — gemini_outcomes deque on ClientState and ProviderState, recent_gemini_outcome_count() method
• Guards screen already parses airlock_gemini and airlock_gemini_response from log records

Approach

Start with percentage display (simple math on existing counters). Drift detection requires a baseline — either a configurable expected ratio or a rolling window comparison. Keep it simple initially: flag when any single mode exceeds 80% of recent responses.