feat(metrics): worker resource dashboard + pprof

[setkyar]

Summary

• Add a standalone, auth-gated worker metrics dashboard for inspecting running pi --mode rpc workers and overall process health — useful for spotting resource hogs, zombie/leaked workers, and diagnosing slowness.
• GET /metrics serves a self-contained HTML page (polls every 2s); GET /api/metrics returns a non-blocking JSON snapshot: process goroutines/heap/SSE-clients/watched-files plus per-worker CPU/RSS via gopsutil.
• CPU% is derived server-side from cumulative-CPU deltas (with a pruned per-PID baseline cache) to avoid gopsutil's blocking Percent(interval). Idle workers alive past the reap TTL are flagged as zombies.
• Mount Go's profiler under GET /api/debug/pprof/ (auth-gated) for deep dives; the dashboard footer links to it.

Design notes

• workers.Manager.Snapshot() returns one WorkerSnapshot per live worker. PID/uptime/idle come from an optional inspector interface implemented by the real rpc worker; test fakes that don't implement it report zeros.
• The OS sampler sits behind a processSampler interface (gopsutil-backed in prod, swappable for tests via Server.SetMetricsSampler).
• pprof.Index hard-codes the /debug/pprof/ prefix, so the index handler is mounted with /api stripped; cmdline/profile/symbol/trace are registered directly.
• gopsutil adds only small, pure-Go (no cgo) deps on macOS/Linux. See docs/dev/metrics-dashboard.md.

Test plan

• make check (vitest + go test + vet + build) green
• Go unit tests: Manager.Snapshot (inspector / non-inspector / empty); /api/metrics JSON shape, auth gating, zombie flag, running-never-zombie, sampler-error degradation, CPU%-delta math, cache pruning; pprof index + named profile + auth gating
• Live smoke (isolated agent dir): /api/metrics valid JSON, /api/debug/pprof/ → 200, /api/debug/pprof/heap → real profile, /metrics footer links to pprof
• Reviewer: open /metrics in a browser with a couple of active chat workers and confirm CPU%/RSS/zombie rendering