fix(visor/ping): narrow ping.mu critical section to map lookup

[0pcom]

Summary

Dominant bottleneck for mux-bw bandwidth measurements. The visor's Ping / PingOnce / PingOnceWithEcho (and dmsg twins) held v.ping.mu — a single visor-global *sync.Mutex — for the entire wire roundtrip (~287ms at 2-hop with 32 KB payloads):

v.ping.mu.Lock()
defer v.ping.mu.Unlock()
pingEntry, ok := v.ping.conns[ref]
// ... ~287ms of wire I/O ...

mux-bw's N pump goroutines each look up their own conn via the map; the wire I/O is independent. But the global mutex serialized them through one ~287ms slot each. So:

• Aggregate throughput across N routes didn't scale with N
• Per-route avg pinned at ~351 kbps even though single-call peak was ~1.7 Mbps (1 RTT × 32 KB)
• --probe-rtt latency probes during a loaded pump measured probe-mutex-wait + network RTT instead of network RTT — swamping the queueing-delay signal at short hop counts (Alpha's lane confirmed this independently)
• Bidirectional simultaneous mux-bw measurements showed mutual-starvation that looked like shared-link contention but was actually mutex contention on each side

Beta and I converged on the same diagnosis independently this afternoon.

Fix

The mutex's actual job is to protect v.ping.conns (the map). Wire I/O on the chosen conn doesn't need the map mutex held — each mux-bw pump goroutine owns its conn via its RouteIndex, no aliasing.

Shrink the critical section to just the map lookup:

v.ping.mu.Lock()
pingEntry, ok := v.ping.conns[ref]
v.ping.mu.Unlock()
if !ok { ... }
// ... wire I/O on pingEntry.conn WITHOUT holding the mutex ...

Applied to Ping, PingOnce, PingOnceWithEcho, DmsgPing, DmsgPingOnce, DmsgPingOnceWithEcho. BandwidthTest already had the correct narrow scope.

Concurrent-close semantics

Pre-fix: StopPing concurrent with PingOnceWithEcho serialized via the mutex — they took turns. Post-fix: StopPing can close the conn while wire I/O is in flight. The Read/Write on the closed conn returns ErrClosed cleanly. mux-bw's pump loop exits on err and surfaces the cause via Beta's MuxRouteFailure event (#2756) so the operator sees the failure mode instead of an indefinite block.

Same-PingRouteRef-from-multiple-goroutines (always undefined behavior on the underlying net.Conn) is unchanged — callers must serialize themselves. mux-bw enforces one goroutine per RouteIndex natively.

Tests

pkg/visor/ping_mu_concurrency_test.go:

• TestPingOnceWithEcho_DoesNotSerializeAcrossRouteIndexes — 200 concurrent calls with distinct PingRouteRefs (no registered conns) complete in ≪ 1s. A regression that reintroduces wire-I/O-under-lock would either time out or take orders of magnitude longer.
• TestPingMu_NotHeldDuringConnAbsentCallpath — after PingOnceWithEcho returns, the mutex must be immediately acquirable from another goroutine. Catches the defer ... Unlock() pattern directly.

Empirical prediction

Once this auto-deploys, the operator's "mux > direct" hypothesis becomes testable without --hops-via intermediate pinning. Per-route avg should rise from ~351 kbps toward the single-call peak of ~1.7 Mbps, and N=2..8 disjoint mux should aggregate roughly linearly (modulo per-intermediate quality variance) instead of flat-lining at single-route throughput.

Chain

The mux-bw measurement loop has now fully closed:

• feat(visor): per-route teardown — decouple StopPing from PK-only #2745 per-route teardown
• feat(router): DisjointMux + ExcludeIntermediatePKs in DialOptions #2746 disjoint-intermediate routing
• fix(visor/router): plumb mux-bw --min-hops through DialPing → router #2749 plumb --min-hops through DialPing
• fix(router): blame intermediate not dst when id_reservation fails on a hop #2750 stop poisoning dst breaker from intermediate failures
• fix(appnet): tryDirectPingDial shortcut bypassed MinHops constraint #2751 tryDirectPingDial gate on MinHops
• feat(visor): honor caller SetupTimeout in DialPing #2752 honor caller SetupTimeout
• feat(rpcgrpc/mux-bw): surface pump-phase route failures as MuxRouteFailure events #2756 MuxRouteFailure pump-phase event
• fix(visor): PingOnceWithEcho adapter must forward RouteIndex + MinHops #2757 PingOnceWithEcho adapter forwards RouteIndex
• this PR: ping.mu doesn't serialize parallel pumps

If the mux-over-intermediates hypothesis is real, it should now be measurable.

Test plan

• go test ./pkg/visor/... -count=1 -short — clean
• go build ./... — clean (via pkg/visor build path)
• go vet ./pkg/visor/... — clean
• gofmt -l pkg/visor/ — clean
• CI lanes (linux, darwin, windows, nix)
• Post-merge: mux-bw N=4 --min-hops 2 should show aggregate avg > 1× single-route avg