todlando — #307 golden #2 (35313912102 on aa8a2978) RED, PRODUCT defect, your lane. RCA: .spt/preserved/307/golden-35313912102/RCA-resume-unbound.md — read it fully first.

DEFECT (one sentence): the brain's periodic liveness reconcile offlines an UNBOUND perch whose CLI-driven resume/run spawn is still in flight, because (a) the CLI path never mints resume custody so `resume_in_flight` cannot defer, and (b) terminal_normalize's CAS is on session_id, which REQ-RESUME-UNBOUND-STAMP preserves. Evidence: a second LIVENESS_RECONCILE_OFFLINE:res1 between SPAWN_PHASE pty_created (2585 ms) and log_ready (4003 ms); perch dump offline carrying run1's session_id.

INVARIANT TO DELIVER: a perch stamped UNBOUND by an in-flight `spt endpoint run|resume` (true-resume AND fresh-fallback branches) is never offlined by reconcile_hosted_liveness before the broker inserts its row; a truly dead harness is still offlined within one tick (every existing reconcile test stays green: pull_liveness_* in livehost.rs, driven_by_selfheal gap_b, attach_wedge_e2e).

FIX SHAPE — your design; two candidates I will accept: (1) the CLI mints custody for its own pid before the spawn so the existing DEFER arm covers it (cleared at spawn return; resume_custody.rs already self-heals a record whose pid is gone); (2) the reconcile's census counts a gate-claimed/in-flight spawn as live (broker KIND_SESSIONS or a sibling query exposing gate_claimed ids). (1) is smaller and reuses a proven guard; if you pick (2), say why. Do NOT widen the CAS to ignore session_id, and do NOT make a time-based grace window the primary guard (a slower box just needs a longer one).

REGISTRY: add a REQ-HAZARD-* to traceable-reqs.toml FIRST (activation is yours), tag impl + int in the same commit; KNOWN-HAZARDS.md entry (Failure / Invariant / spt-core mapping) citing this RCA.

PROOF I need before the gate: a deterministic repro test that FAILS on main 0ab1e48b and PASSES with the fix — e.g. an UNBOUND perch + empty live set + the in-flight marker → reconcile must DEFER; the same perch WITHOUT the marker → still offlined (negative control); plus one real-broker int leg if the fix touches the census. Then clippy --workspace, traceable-reqs check, targeted nextest of rc_attach_truth (all 6) + the reconcile suites. Branch off current main (0ab1e48b), thin PR, send me the sha; I run the isolated second-producer gate before ff.

Pool: claim your own worktree pool from inside it (pool-claim --label). Box: golden #2's remaining Windows jobs may still be running on the hfenduleam runner — check `gh run view 35313912102 --repo BigscreenVR/spt-bs-core --json jobs` before any cargo; sequence around Runner.Worker, never relocate it.