fix: prevent false spawn failures for live agents

[ruby-dlee]

Summary

Fixes six observed false spawn failures on 2026-07-29 across the Firstmate and Relvino pools, including this task's own Codex spawn. In every incident the worktree-path readiness probe timed out even though the harness agent was provably alive and working.

The root cause was shared with fm-herdr-send-fix-h7: Herdr path reads were gated by a legacy reachability/certificate path that could reject a healthy native server. This branch builds on that fix rather than duplicating it.

What changed

• After the 60-second path-readiness loop, fm-spawn.sh now independently asks the backend's conservative harness-agent liveness classifier before condemning an unreadable endpoint.
• An unreadable path plus confidently alive commits the spawn through the normal success path, with no direct_spawn_cleanup, direct_spawn_artifacts, or rollback_pending poison markers.
• A dead or unknown agent still enters the existing rollback path.
• A concrete wrong-worktree observation remains a hard failure even when an agent is alive.
• Regression coverage exercises all three outcomes.
• No-mistakes also hardened secondmate test fixtures so default-branch resolution works without assuming a local main branch in CI-created clones.

Verification

Implementation-specific automated coverage is green:

• The new readiness regression failed before the fix with the reproduced 60-poll timeout.
• FM_TEST_FOCUSED=spawn-readiness bash tests/fm-account-directory.test.sh passes the live-agent success, genuinely dead rollback, and wrong-worktree rollback cases.
• The full tests/fm-account-directory.test.sh affected suite passes.
• FM_TEST_FOCUSED=readsteer-native-cutover bash tests/fm-backend-herdr.test.sh passes the coordinated Herdr read/steer regressions.
• Repository-pinned ShellCheck lint and Agent Fleet tests pass.
• GitHub reports 4 passed, 0 failed, and 1 skipped check. Behavior Tests was cancelled and recorded as skipped while Firstmate reduced machine saturation; no-mistakes declared the PR CI-ready.

Live validation limitation

No live end-to-end Herdr spawn was run because this task was not scaffolded with --herdr-lab, and its safety contract forbids touching Herdr lifecycle without that guard.

That leaves real backend timing, native process registration, and the post-land behavior of an ordinary production spawn unproven by this branch. The real-world reproduction is unusually strong—six independent incidents across two pools, all with active Codex agents—but it is reproduction evidence, not post-fix live validation.

Firstmate will verify the next ordinary spawn after landing and report whether it records clean success metadata. An isolated Herdr-lab lifecycle regression is being filed separately.

Risk

The fallback is deliberately narrow: only confident harness-agent liveness can rescue an unreadable readiness probe. Concrete isolation violations and dead or unknown endpoints still fail closed and roll back.