tools/stress: agent SSH + log parser

[elitegreg]

Summary

Completes the device-stress orchestrator (#3746) by replacing the no-op AgentRunner with the live SSH-driven runner and the log parser that turns agent diff/commit lines into pre_commit_log / applied runlog rows. Stacked on top of #3776 (part 2, orchestrator skeleton). Part 3 of #3746. Closes #3772.

• pkg/agent/parser.go — Parser.Parse(line) []Event tracks two log lines from controlplane/agent/pkg/arista/eapi.go:
  • Committing config session due to diffs detected: <diff> → extracts every + interface Tunnel<ID> and emits one pre_commit_log event per ID; the diff's - lines (deprovisions) are ignored.
  • Configuration session finalized with command '... commit' → emits one applied event per pending tunnel; the ... abort variant clears the buffer without emitting.
• pkg/agent/ssh.go — NewSSH(cfg) Runner dials --dut-ssh-host with --dut-ssh-key, execs doublezero-agent -verbose (appending -controller <addr> when --controller is set), and tees remote stdout/stderr into <working-dir>/orchestrator.agent.log while running every line through Parser. The session is closed on ctx cancel; the events channel closes after both stream readers exit so consumers never see a half-emitted event. Host-key verification uses ssh.InsecureIgnoreHostKey (documented; targets are ephemeral cEOS containers).
• pkg/sweep — gained an agent-event consumer goroutine and a tunnelRegistry populated as users are created; consumer attributes each agent.Event back to a user_index via tunnel ID lookup and appends pre_commit_log / applied rows. Unknown tunnels are debug-logged and dropped. The agent is started under a derived context that the sweep cancels after deprovision, so a clean run still shuts the agent down rather than leaking goroutines.
• pkg/exec.fetchTunnelID — now implemented: GetAccountInfo on the user PDA + DeserializeUser + return TunnelId. exec.Config gains an RPC field; the cmd binary passes the same *solanarpc.Client used by Client/Executor.
• cmd/device-orchestrator — new flags --dut-ssh-user (default admin) and --no-agent (offline testing). SSH runner is the default when --dut-ssh-host and --dut-ssh-key are both set; with either missing, the cmd falls back to the no-op runner and warns.

Testing Verification

• pkg/agent/parser_test.go: golden line fixtures for single-tunnel diff, multi-tunnel diff (mixed +/-), deprovision-only diff, commit-success after multi-tunnel diff, abort-clears-buffer, stray commit-with-no-pending, two consecutive provision cycles, oversized tunnel ID skipped, Tunnel5000 vs Tunnel500 boundary.
• pkg/sweep/sweep_test.go: new scriptedAgent + a deleteGate on the fake executor lets the test emit agent events while deprovision is blocked; asserts pre_commit_log / applied rows are written for the two registered tunnels and the unregistered tunnel 999 is dropped. Tests pass under -race.
• pkg/exec/exec_test.go: stub RPC returns a hand-encoded User body with TunnelId = 4242; fetchTunnelID reads it correctly. Missing-account path returns an error containing not found.
• Smoke test: make build produces bin/device-orchestrator; --dry-run writes orchestrator-config.json containing dut_ssh_user, no_agent, and the rest of the flag set.
• make go-build go-lint go-test all green.

Out of scope

• Integration test that actually SSH-dials a container. The SSH client is deterministic plumbing over golang.org/x/crypto/ssh; CI never opens an SSH session. Acceptance is via the manual devnet run per stress: implement tools/stress/device-orchestrator #3746.