workflow-fix: GCP gate-park teardown leg + terminal-phase reconnect refusal

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from the orchestrator's own incident on task #763 (cofit round, 2026-07-03).

Goal

Add a GCP-lane instance-teardown leg to the /issue Step 6d.4 gate handlers (and the gate-handler authoring guidance): on a BLOCKING gate emitted by a GCP-lane workload, the orchestrator deletes the instance after the sentinel is drained and the gate is resolved — and the router's reconnect path must never treat a gate-parked/terminal-phase instance as a live reconnect target.

Workflow gap

• Bug observed (two legs, one incident): (1) #763's cofit_phaseA_done gate-park kept the GCE instance RUNNING after [phase=done] — correct for sentinel draining, but no handler step tears it down; the orchestrator resolved the gate (launched the off-pod phase) without deleting the instance, and an A100-80 idled ~40 min. (2) The NEXT dispatch_issue.py launch for the same issue returned reason: reconnect to that zombie (RUNNING, terminal eps/phase=workload_done) WITHOUT launching the new workload — silently no-op'ing the phase-C dispatch until the orchestrator noticed the poll reading workload_done.
• Why it is a workflow gap: Step 6d.4 documents the RunPod pod-cycling teardown (pod.py stop) for pv_phase1_done but has no GCP-lane analogue; gcp.reconnect_or_none treats any RUNNING eps-issue-<N> as reconnectable without checking the workload phase is non-terminal. Both legs will recur for every multi-phase GCP-lane experiment that gates mid-pipeline.
• Confidence (emitter): high (reproduced live on #763; markers 03:41-03:52Z).

Proposed change (candidate diff sketch — refine in planning)

.claude/skills/issue/SKILL.md, Step 6d.4:
+ GCP-lane blocking gates: after the poller drains the sentinel and the
+ handler resolves the gate (parks OR auto-resolves), if the handle's
+ backend is gcp and the workload printed [phase=done]/a terminal gate
+ park, DELETE the instance (gcloud compute instances delete, or
+ dispatch_issue.py finalize --skip-confirm-artifacts when mid-pipeline)
+ BEFORE dispatching the off-lane phase — the instance stays up ONLY for
+ sentinel draining, never through an off-pod phase (8-bis).

src/explore_persona_space/backends/gcp.py reconnect_or_none:
+ refuse to reconnect to an instance whose eps/phase guest attribute is
+ terminal (workload_done / a blocking-gate park) — return None so the
+ router creates a fresh instance.

Scope / surfaces

• Primary targets: .claude/skills/issue/SKILL.md (Step 6d.4), src/explore_persona_space/backends/gcp.py (reconnect predicate — in-scope backends module).
• Grep: grep -rn "reconnect" src/explore_persona_space/backends/gcp.py scripts/backend_poll.py; check the janitor's zombie predicate (gcp_audit.py) for the same RUNNING-but-terminal-phase signature to share it.

Constraints / invariants

• Workflow-surface only; router tests (tests/test_router*.py, tests/test_gcp_backend.py) stay green; add a pin for the terminal-phase reconnect refusal.

Provenance

• workflow_fix_target: .claude/skills/issue/SKILL.md, src/explore_persona_space/backends/gcp.py
• fingerprint: 3cbb26c26d52