workflow-fix: RunPod lane drops --workload-cmd (provision-only launch)

Overview / Motivation

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

Goal

Make RunPodBackend.launch actually EXECUTE a --workload-cmd (nohup on the pod + pod-side pid file + log at the handle's log_path + a run-launched sidecar/marker), instead of provision-and-bootstrap-only.

Workflow gap

• Bug observed: dispatch_issue.py launch --issue 763 --backend runpod --workload-cmd 'bash scripts/issue763_cofit_dispatch.sh --phase-group C' returned ok (pod-763 provisioned, bootstrapped, handle sidecar written with log_path=/workspace/logs/issue-763.log) but the workload NEVER started: no process, empty /workspace/logs, first poll tick read status=dead/unknown. The orchestrator had to SSH in and nohup the command manually (+ post epm:run-launched by hand).
• Why it is a workflow gap: the #588 contract ("every lane executes custom dispatch scripts via --workload-cmd") does not hold on the RunPod lane — the launch path silently drops the workload command, so every RunPod --workload-cmd failover (the GCP-to-RunPod doctrine's own target lane) lands as a healthy-looking pod running nothing, indistinguishable from a crashed run to the poller.
• Confidence (emitter): high (reproduced live; the manual SSH launch of the identical command ran immediately).

Proposed change (candidate diff sketch — refine in planning)

src/explore_persona_space/backends/runpod.py (launch):
+ when spec.workload_cmd is set: after bootstrap, ssh-exec a launch leg that
+ (1) syncs the pod clone to spec.repo_branch (fetch + reset the pod-side
+     tree to origin/<repo_branch>; pod-side only, never the VM tree),
+ (2) sources the pod .env, ensures /workspace/logs exists,
+ (3) nohup-runs the workload_cmd redirecting to <log_path>, writes the
+     pid to a pod-side pid file,
+ (4) verifies the log file exists before returning, and records pid +
+     log_path in the RunHandle sidecar (mirroring the experimenter contract).
scripts/dispatch_issue.py: surface a LOUD error / nonzero exit if a runpod
--workload-cmd launch returns with no live workload pid (fail-fast instead
of ok:true).

Scope / surfaces

• Primary targets: src/explore_persona_space/backends/runpod.py (in-scope backends module), scripts/dispatch_issue.py.
• Grep: grep -n "workload_cmd" src/explore_persona_space/backends/*.py scripts/dispatch_issue.py scripts/pod_lifecycle.py — confirm which layer drops it; add a router test pinning the RunPod workload-cmd execution (tests/test_router*.py / tests/test_backend_*.py).

Constraints / invariants

• Workflow-surface only; the pod NEVER shells scripts/task.py (sentinel/sidecar signaling only); router tests stay green.

Provenance

• workflow_fix_target: src/explore_persona_space/backends/runpod.py
• fingerprint: be5f3fa75f02