feat(usage): token collection pipeline + rate-limit bucket tracking

[fdatoo]

Summary

• Token collection per run: A usage_tailer actor starts alongside each container, polls runs/<id>/mount/**/*.jsonl every 2s, parses Claude's JSONL output, and atomically writes token counts to run_usage_events + runs.tokens_*. Live {"type":"usage"} frames flow over the run WebSocket so RunUsage updates in real-time during a run.
• Rate-limit buckets + Usage page: Rate-limit headers embedded in JSONL lines are parsed into rate_limit_buckets. GET /api/usage now returns real data (utilization bars, countdown timers) instead of the missing_credentials stub. The usage WebSocket publishes {"type":"snapshot"} frames on every bucket update.
• Threshold notifications + pacing: Crossing 75% or 90% emits {"type":"threshold_crossed"} on the usage pubsub (picked up by UsageNotifier for browser notifications). Pacing verdicts (chill/on_track/hot) are computed from window_started_at/reset_at and included in every snapshot.

Test plan

• Start a run and watch RunUsage in the drawer populate with live token counts
• Check /usage page shows bucket utilization bars and countdown timers after the first JSONL line with rateLimits is parsed
• Verify GET /api/usage/daily returns non-zero token totals after a completed run
• Confirm browser notification fires when a bucket crosses 75% (requires notifications enabled in Settings)
• make test-server passes (48 tests)

🤖 Generated with Claude Code