Surface first-token gate + confidence-suppression counters in Settings/Diagnostics

[Jam-Cai]

Problem

Both first-token gates emit to Console.app only:

• Deny-list gate (Logit gating on first generated token #24, PR feat: First-token logit gating to suppress chat-residue openers #84) → `subsystem == "app.tabby" AND category == "first-token-gate"`
• Confidence gate (Confidence-based suggestion suppression #81, PR Suppress low-confidence first-token suggestions (#81) #88) → `subsystem == "app.tabby" AND category == "first-token-confidence"`

To see "how often does this fire?" a user has to attach Console.app and run `log stream`. That's fine for development; it's a bad feedback loop for users tuning their threshold or for contributors verifying the gates are doing anything.

Proposal

Add a small "Diagnostics" section (or extend the existing autocomplete settings) that displays:

• Suggestions generated this session
• Suppressed by deny-list gate
• Suppressed by confidence gate
• (Optional) a rolling histogram or last-N display of first-token top-1 probabilities

Counters reset on app launch — no persistence, no PII.

Acceptance

• A counter type on `LlamaSuggestionEngine` (or coordinator) that increments on the two suppression paths and on successful generations.
• A simple SwiftUI view exposing those counters live.
• No new logs; this is purely a UI surface over existing signal.

Out of scope

Cross-session persistence, opt-in upload, anything that turns this into telemetry. That belongs with #27 (opt-in encrypted capture), not here.