fix: a failing pprof listener must not kill the media server (#64 item 1)

[vsolano9]

Item 1 of #64. Independent of the rest of that list; items 3–10 are untouched.

The defect

startDebugServer's goroutine ended in log.Fatalf, which is os.Exit(1):

if err := srv.Serve(listener); err != nil && err != http.ErrServerClosed {
    log.Fatalf("debug server error: %v", err)
}

That skips sm.CloseAll() and both graceful shutdowns in main, so an accept error on an optional profiling socket drops every live WebRTC call with no cleanup. As the issue puts it, this mirrors the main HTTP server's pattern in a place where the reasoning does not carry: the process genuinely cannot do its job without the media listener. It can without pprof.

Now log.Printf, and the server keeps running without profiling.

Making it testable

The goroutine literal is lifted into a named serveDebug(srv, listener) so its return is observable. Nothing else moved.

Two tests, because the obvious one is not sufficient on its own:

• TestServeDebugSurvivesAFailedListener drives serveDebug with a listener whose Accept returns a permanent error, and requires the call to return. The error is deliberately not a net.Error with Temporary() == true — net/http retries those with a backoff, so a temporary error would hang rather than exercise the exit path.
• TestPublicMuxStillServesAfterTheDebugListenerDies goes through the real startDebugServer, stops the debug listener under its own Serve loop, then requires the public mux to still answer /health. On its own the first test would be satisfied by a serveDebug that nothing calls; this one asserts the thing an operator actually cares about.

Red proof

Reverting only the log.Printf back to log.Fatalf, with both tests in place:

=== RUN   TestServeDebugSurvivesAFailedListener
FAIL	github.com/streamcoreai/streamcore-server	0.355s
FAIL

Worth reading closely: there is no --- FAIL: line. The test does not report a failure — the test binary is killed mid-run, which is the defect reproduced inside the suite. A reader who greps for --- FAIL would conclude nothing ran.

Validation

The exact CI commands from .github/workflows/ci.yml:

• gofmt -l . — clean
• go build ./... — clean
• go vet ./... — clean
• go test -race ./... — ok across all 15 packages

Not claimed

I have not reproduced a real accept failure on a live socket; the listener is a stub, which is what makes the test hermetic. And this changes only the crash behaviour — a failed pprof listener now leaves the server running without profiling, silently apart from the log line. If you would rather it also surface in /health or a metric, that is a design call I did not want to presume, and it belongs with item 7 (warning when profiling settings cannot take effect) rather than here.

[jason-shen]

looks good to me, thanks @vsolano9