SwiftFTR 0.15.0

This release makes traceroute survive two conditions that previously destroyed a measurement:
momentary send-buffer pressure, and a system resolver that stops answering. Both were found in
downstream production telemetry rather than in testing.

Traces no longer fail on transient send pressure

Probe sockets are non-blocking and a trace pushes every hop's probe out in one burst, so a single
sendto returning EAGAIN/EWOULDBLOCK/ENOBUFS — the kernel saying "no room right now" — used
to fail the entire traceroute with TracerouteError.sendFailed. This was the most common traceroute
failure observed in the field.

Those three errnos are now retried: poll(POLLOUT) for socket-buffer pressure, and a 1 ms sleep for
ENOBUFS, which reports the interface output queue and leaves the socket reporting itself writable.
The budget is 250 ms shared across the whole burst rather than per probe, so the default 1 s receive
window is not at risk. Ping applies the same retry with a 50 ms per-send budget, so transient
pressure no longer silently drops a packet from the sample.

Every other errno still fails on the first attempt with no added latency. Callers that map
EHOSTUNREACH, ENETDOWN, ENETUNREACH and EACCES to offline or permission states see them
exactly as promptly as before. When the retry budget is exhausted, the same sendFailed(errno:) is
thrown as in 0.14.

traceClassified() is now bounded and cancellable

traceClassified() completes, or throws, within a budget derived from its own configuration, and
honors task cancellation during enrichment as it already did during the probe phase. A cancelled
classified trace now returns in about 0.1s, where it previously returned only when a slot on the
shared blocking-I/O executor happened to free.

The enrichment phase could previously run far past any caller watchdog while ignoring cancellation.
getaddrinfo (once per STUN hostname, walked serially) and getnameinfo (once per hop) have no
timeout of their own and stall 30 seconds against a resolver that drops queries silently, so
public-IP discovery alone could consume 90 seconds before a single hostname lookup began.

Reverse DNS now degrades rather than stalling — read this if you rely on hostnames

This is the one change that can take something away. Reverse DNS gives up after
rdnsLookupTimeout (default 5s), and two consecutive stalled lookups suppress further attempts
until the cache is cleared, which callers already do on a network change. Hops whose lookup does not
finish in time report numerically.

Previously a slow-but-working resolver would eventually return a hostname — after up to 30 seconds.
On genuinely slow links, hostnames that used to appear may now be absent. If you see fewer
hostnames after upgrading, raise SwiftFTRConfig.rdnsLookupTimeout; the budget bounds the
reverse-DNS phase as a whole rather than being paid per hop, because lookups run concurrently, so
raising it costs one wall-clock wait and not one per hop.

Public-IP discovery degrades the same way: classification proceeds without a public address when
discovery exceeds its budget.

New public API

• SwiftFTRConfig.rdnsLookupTimeout and SwiftFTRConfig.publicIPDiscoveryTimeout set the
  enrichment budgets. Both are optional and fall back to the defaults when absent or invalid,
  matching the existing rdnsCacheTTL / rdnsCacheSize pattern.
• SwiftFTRConfig.defaultRDNSLookupTimeout (5s) and SwiftFTRConfig.defaultPublicIPDiscoveryTimeout
  (6s) expose those defaults, so callers can adjust relative to them.
• TraceClassifier.classify(...) gains a defaulted publicIPDiscoveryTimeout parameter for callers
  that reach discovery without a config in hand.

All additions are additive and defaulted; existing call sites compile unchanged.

Known limitation

A blocking syscall that has already started still occupies its executor slot until it returns — up
to 30 seconds for a stalled resolver. The calling trace is released; the slot is not. Removing this
requires not calling getnameinfo at all, which depends on discovering the effective system
resolver. Tracked in ROADMAP.md.

What's Changed

• docs: polish 0.14 public guidance by @dweekly in #55
• docs: close out 0.14 release by @dweekly in #56
• docs: plan two production traceroute failures by @dweekly in #57
• fix(enrichment): bound and cancel the enrichment phase by @dweekly in #59
• fix(tracer): retry transient sendto pressure instead of failing the trace by @dweekly in #58
• docs: drop shipped roadmap items by @dweekly in #60
• chore: prepare 0.15.0 by @dweekly in #61

Full Changelog: v0.14.0...v0.15.0