fix(profiler): keep partial results when a PID fails, and explain pprof failures

[mayankpande88]

Summary

• A PID the profiler can't attach to no longer fails the whole run. That can be a shell, a helper process, or a process that isn't the target runtime.

  • Every PID is profiled, and each one that succeeds publishes its result as before.
  • The run succeeds when at least one PID did, and emits a notice listing the skipped PIDs and why.
  • It fails only when every PID failed, with each reason: PID <n>: <reason>; PID <m>: <reason>.

• A flamegraph that can't be rendered now fails that PID (bpf, perf, py-spy, austin). Before, the run ended as a success with no result.

• Go pprof scrape failures name the cause:

  • the endpoint requires authentication (401/403);
  • no /debug/pprof handler (404);
  • nothing listening (connection refused).

  When the listening port can't be detected, the error says the default :8080 was tried.

Type of change

• Bug fix
• CI / build

Test plan

• Unit tests:
  • the shared helper: mixed results, all failing, all succeeding, the notice, stagger and concurrency;
  • each profiler's all-PIDs-fail path;
  • the flamegraph-failure path for the four tools;
  • pprof error mapping and the port fallback.
• New Docker e2e scripts, run in Code Verify:
  • test/e2e/python-austin-multi-pid.sh: a Python process next to a non-Python one, both start orders, with the PIDs found by the agent itself.
  • test/e2e/go-pprof.sh: Go servers with pprof open, behind basic auth, not registered, and with nothing listening.

Engineering detail

• Shared helper. common.ProfilePIDs replaces eight copies of the worker-pool fan-out.
  • It keeps the stagger between PIDs, but no longer sleeps after the last one.
  • An empty PID list is now an error.
  • The alitto/pond dependency is dropped.
• Result events are unchanged: one result event per published file. A partial success still emits the successful PID's result first, followed by the notice and progress: ended. Before, a failing PID ended the run with an error event, after any result already published, and PIDs not yet started were cancelled.
• Go pprof:
  • Each PID gets its own copy of the job. Concurrent tasks were setting PID on a shared one.
  • The scrape runs through the commander, and failures are read from busybox wget's stderr.
• The test fakes now lock in invoke, which -race needs once PIDs run on separate goroutines.

Local e2e: built the python and bpf profiler images from this branch (arm64) and ran test/e2e/python-austin-multi-pid.sh, test/e2e/go-pprof.sh and test/e2e/python-austin.sh; all passed.

• Multi-PID: in both start orders, the Python PID's profile was published with samples (about 400 each), and the non-Python PID was named in a notice: Profiled 1 of 2 PIDs; skipped PID 8: could not launch profiler: austin (PID 8, /usr/bin/sleep): ….
• Go pprof: the open pprof server produced a profile. The auth, no-handler and no-listener servers produced the pprof endpoint on :6060 requires authentication, no /debug/pprof handler on :9090 (is net/http/pprof registered?) and could not detect the listening port (…), so tried the default :8080: … nothing listening on :8080.
• Negative control: against images built from main, both new scripts fail. The multi-process run ends with an error event and no notice, and the pprof errors are raw wget output.

Checks:

• go build, go vet and go test -race ./internal/agent/profiler/... are clean on Linux.
• golangci-lint shows no new findings, and fixes one existing QF1003.

Checklist

• go build ./... and go test ./... pass locally
• Docs (README, flags) updated if behavior changed: no flag or doc change
• No breaking changes to the agent's stdout JSON envelope. Partial success now also emits a notice event, which is an existing event type.

[gemini-code-assist]

Code Review

This pull request refactors concurrent profiling across all supported languages and tools by replacing the third-party pond library with a new custom common.ProfilePIDs helper. This helper runs profiling concurrently with a staggered delay and handles partial failures gracefully, ensuring that a single failing PID does not fail the entire run. Additionally, error reporting has been improved (especially for Go pprof scrape failures), fake managers have been updated with mutexes to serialize concurrent invocations in tests, and new E2E tests have been introduced. However, a critical compilation error was identified in the new common.ProfilePIDs implementation, where a non-existent Go method is called on a standard sync.WaitGroup.