feat: wave 2 — observability (telemetry, SSE logs, /metrics)

[wra-sol]

Summary

Wave 2 of the opencode manager improvement plan. Three slices, one PR. Based on Wave 1 (PR #19).

2.1 — Expose crash count, uptime, last-exit rc, restart history (closes #6)

ChildProcess now tracks started_at, last_exit_rc, last_exit_at, crash_count, and a restart_history deque (last 10 entries with timestamp, reason, and rc). These are surfaced in:

• /manage/status JSON (new fields: pid, started_at, uptime_s, last_exit_rc, last_exit_at, crash_count, restarts)
• Dashboard HTML (new rows: Uptime, Crashes, Last exit)

Restart reasons tracked: manual, crash, reconfigure, stopped.

2.2 — SSE log tail + manager events in ring + filter/search (closes #7)

• Log ring broadened from 400 to 1000 entries, now captures both child stdout (plain strings) and manager events (dicts with ts, source, level, line).
• Manager events logged: login success/failure (with IP), restart, key rotate, reconfigure, proxy errors, child crash/respawn/give-up.
• New endpoint GET /manage/logs/stream — SSE stream that tails the log ring live. Sends the current filtered ring as an initial burst, then appends new entries as they arrive (polls at 0.5s).
• Filter params on both /manage/logs and /manage/logs/stream: ?level=error, ?source=child or ?source=manager, ?q=<substring>.
• /manage/logs returns structured entries (child stdout normalized to dicts for the API).
• Dashboard renders both string and dict log entries.

2.3 — /metrics Prometheus endpoint (closes #8)

New endpoint GET /manage/metrics (session-auth required) emitting Prometheus text format:

Counters (monotonic, process-lifetime):

• opencode_railway_requests_total{path}
• opencode_railway_login_attempts_total
• opencode_railway_login_failures_total
• opencode_railway_child_crashes_total
• opencode_railway_restarts_total{reason}
• opencode_railway_proxy_errors_total

Gauges:

• opencode_railway_child_up (1|0)
• opencode_railway_log_ring_size
• opencode_railway_uptime_seconds

Metrics class with thread-safe increments. Request counting in do_GET and do_POST.

Tests

11 new test cases:

• Status telemetry fields, restart history recording, dashboard telemetry display
• Log filtering by source, search, manager events in ring, structured log entries
• /metrics requires auth, returns Prometheus text, counts requests, counts login failures

119 total tests pass. ruff check + ruff format --check clean.