test(auto-resume): gate the crash-once claude shim on the create's settled reply

[danshapiro]

What

Fixes the second base-gate flake at origin/main: crates/freshell-ws/tests/auto_resume_e2e.rs > successful_auto_resume_respawn_does_not_drop_its_ticket_unarmed timed out waiting for a terminal.created frame (~1-in-7 file runs under cargo test --workspace parallel load; 0/15 isolated; 0/12 single-threaded; reproduced locally and under the gate).

RCA (probe-instrumented, receipts in the investigation)

The claude crash-once shim died within milliseconds of the PTY spawn — racing the create's own post-spawn tail. When the death won:

1. The b8ke ext r9 dead-PTY guard (commit_session_ref_ownership: "a dead runtime NEVER records Live") refused the create's winner commit (ForeignOperation — confirmed by probe, not from the coordinator's state match).
2. The create answered the designed typed error error{INTERNAL_ERROR: lost session ownership during spawn} (the Err commit arm's stale-teardown).
3. Gate 1 worked exactly as designed — the hub's crash event was held (pending=true) until the create answered, then the hub claimed and respawned.
4. The test only accepts terminal.created, silently consumed the error frame, and burned its whole 30s frame budget.

So the production flow is the designed outcome for "child dies during create settle"; the test was racing its own deliberate crash.

Fix

The shim's first generation now waits on a per-test release marker (50ms poll) that the tests write AFTER terminal.created — the crash is structurally post-settle, with no timing assumption anywhere (the codex sibling's fixed 1.2s sleep is the cruder wall-clock-margin precedent). Both crash_once_claude_spec call sites updated.

Evidence

• Isolated: 0/15 failed. Single-threaded: 0/12. Parallel pre-fix: ~12% failure. Post-fix parallel: 0/35 failed.
• cargo fmt --check clean; pre-push gate green.