Phase 4 of the firehose rework. Adds the active relay-health prober — the piece that refreshes the relays directory the /relays page renders (data is currently a frozen 2025-11 snapshot). Unlike the subscription hoses it is active (dials each relay) and kind-less, so it is not a hose — it's its own self-scheduling module/process.

Design (confirmed)

• Process model: src/prober.js — sweeps on an interval by default, runnable standalone or under pm2 (probe.js entry, mirroring serve.js), with --once / RUN_ONCE for a single sweep.
• Check depth: WebSocket connect (online, responseTime, uptime rollup, lastChecked, lastError) plus a NIP-11 HTTP GET (Accept: application/nostr+json) for requiresAuth / requiresPayment. acceptsEvents stays untouched (publish-test is a later pass; the page already renders missing as "—").
• Relay universe: the relays collection itself — continuously self-fed by the phase 2/3 harvesting.
• Consistent rollup: aggregation-pipeline updateOne so uptime = checksOnline/checksTotal*100 is computed from the freshly $inc'd running totals.
• Scale knobs (env, CLI-overridable): PROBE_INTERVAL (default 1h), PROBE_CONCURRENCY (25), PROBE_TIMEOUT (7s).
• No new deps — Node's global WebSocket / fetch.

Scope

• src/prober.js: pure helpers (proberConfig, nip11Flags, relayInfoUrl, arg parsing) + checkRelay, fetchNip11, probeRelay, sweepOnce, runProber.
• probe.js: pm2/standalone entry; npm run probe.
• .env.example + README: document the prober and its env/flags.
• Tests: pure helpers (config/flags, NIP-11 flag extraction, ws→http URL, scalar handling).

Out of scope

Publish-test (acceptsEvents) — a later pass · deploy/cut-over on nostr.social — phase 5.

Acceptance

• node probe.js --once sweeps the relays collection and updates online/responseTime/uptime/lastChecked (+ requiresAuth/requiresPayment from NIP-11) without clobbering unrelated fields.
• Self-schedules by default; --once exits after one sweep.
• npm test green incl. new prober unit tests.