feat: outage measurement, TUI path trace, and history sparkline

[servak]

Summary

1. Outage measurement (for failover / maintenance tests)

• An outage starts after 3 consecutive lost probes and ends at the next successful probe. It is recorded with start/end, duration, and lost probe count; the start is the send time of the first lost probe. Shorter runs count as ordinary packet loss. The threshold is fixed, not configurable.
• Out-of-order results: concurrent TCP/HTTP probes can report a timeout after a later probe has already succeeded. Results are therefore buffered for timeout + interval and applied in send-time order. The buffer is flushed when the event channel closes, so final reports are complete.
• Where it shows up:
  • Host list: the LastFailTime cell also shows the outage duration, e.g. 16:44:20 (DOWN 5.20s) in red while down and 16:44:15 (0.50s) in yellow after recovery.
  • TUI detail panel: outage count, total downtime, longest outage, and the most recent outages.
  • A cross-target timeline printed on TUI exit and in batch table output.
  • New json/csv fields: outages[], outage_count, total_downtime_ms, longest_outage_ms.

2. MTR-style path trace in the TUI

• New internal/trace package. Each round it sends ICMP echo with TTL 1..30 and matches replies by echo ID/seq, including the datagram quoted in Time Exceeded / Dest Unreachable messages (handles IPv4 IHL options and IPv6).
• Uses a raw socket when available and falls back to unprivileged datagram ICMP. On macOS the fallback also receives Time Exceeded, so it works without root (verified).
• p toggles a live path panel that follows the selected host. Any target type works: for tcp://, https://, dns://, … the host part is traced.
  • The trace starts asynchronously and runs at most one round per second, so router ICMP rate limiting doesn't show up as fake loss.
  • Probes within a round are spread evenly over the interval (like mtr). While the destination is unknown, it probes at most 8 TTLs past the farthest hop that has replied. Sending TTL 1..30 back to back made the home router treat the traffic as an ICMP flood and drop ICMP for every host behind it, which also broke mping running on other machines.
  • Hops still waiting for their first reply show waiting. The panel also notes when loss appears only at intermediate hops, which is usually ICMP rate limiting rather than real loss.
• There is no standalone mping mtr subcommand; the trace is only available in the TUI.

3. History column and percentiles

• The host list shows the last 20 probes as a sparkline, and × marks a failure. Bar height is how many milliseconds an RTT exceeds the host's median (+5/10/20/50/100/200/500ms). A host's normal jitter therefore stays flat, and +100ms or more is drawn in the warning color.
• RTT p50/p95/p99 in the detail panel and in json/csv.

Fixes

• Data race between probing and the UI: SortBy/GetMetrics returned live pointers that the UI read without the lock. They now return deep-copied snapshots, and panels are refreshed from the latest snapshot on every update.
• Selection: the selection now stays on the same host when re-sorting moves it to another row, and returns to that host after a filter that hid it is cleared.
• Descending sort: it uses a strict ordering, so equal rows no longer swap places on every refresh.

Cleanup

• The failure beep is disabled through stats.Options.DisableBeep. This replaces SetBeepEnabled, which was added in feat(batch): add json/csv output, max-loss exit code, and RTT jitter #58.
• NewMetricsManagerWithHistorySize is removed; use NewMetricsManagerWithOptions.

Test plan

• go test -race ./..., golangci-lint run (0 issues)
• Unit tests:
  • Outage detection: 3-probe threshold, short loss ignored, ongoing outages, reordering, late results.
  • Exports, sparkline, percentiles, LastFailTime cell.
  • Selection across re-sort and filter.
  • ICMP reply parsing (v4/v6, IHL options, rejects), target → host extraction, rate-limit detection.
• Network integration tests for path tracing (-tags integration)
• E2E: stopped and restarted a local server during batch → 1.600s outage = 8 lost × 200ms
• TUI in tmux: History column, outage detail, path view following the selection
• Path view without root on macOS (11 hops, reverse DNS)
• Paced probing: 8.8.8.8 and 1.1.1.1 traced in parallel reach the destination in 15/15 rounds, with 0/200 loss to the gateway meanwhile (the burst version never reached either destination and caused shared loss on the LAN)
• Cross-compiled for every goreleaser target (linux/windows/darwin × amd64/arm64/386)
• Raw socket on macOS (setuid root): ICMP monitoring and the TUI path view in the same process with no cross-talk (29/29 echo replies). The ICMP prober ignores the tracer's Time Exceeded messages by echo ID.
• Raw socket (root) run on Linux
• IPv6 path trace on a v6-enabled network (no v6 route here; parsing is covered by unit tests)

🤖 Generated with Claude Code