EPS Dashboard

Overview / Motivation

Auto-filed by the /daily Step-C parked-candidate sweep (2026-07-08) from a candidate parked on task #1087 by a recursion-guarded workflow-fix session.

Goal

Mechanize the realized-keys check: a pre-provision key-schema probe (torch.load(path, map_location='cpu', mmap=True).keys() superset-of the consumer's asserted keys) in the reuse-staging path, or a pytest running the consumer loader against pinned bundles.

Workflow gap

Bug observed: The realized-keys check landed as rule prose via #1087 (artifact-reuse.md check (c)/(h)(ii): torch.load(mmap=True).keys() / consumer-loader-open, commit b61d3e623f), but nothing runs it mechanically — a planner/implementer who skips the prose check still stages a bundle whose realized key set misses the consumer's asserts (origin incident #1073).
Why it is a workflow gap: Prose-only fitness checks in artifact-reuse.md depend on agent compliance; both #1073 sweep sites evaded the manual check, and the critics that reviewed them missed it too — the same class the staged-layout consumer-open probe (#928) mechanized for layout.
Confidence (emitter): medium
Sweep verification (2026-07-08): Rule-level fix confirmed present (artifact-reuse.md lines 116-117, 181 — landed by #1087 itself, commit b61d3e623f); no mmap key-schema probe exists in scripts/ or the backends (grep 2026-07-08); open wf-fix #865 on scripts/ is unrelated (Step 9c selector worktree blindness). The MECHANIZATION leg is the unimplemented residue.

Proposed change (candidate diff sketch — refine in planning)

(none — synthesized from prose follow-up; likely shape: a helper in the reuse-staging path that torch.loads each staged bundle with mmap=True and asserts .keys() is a superset of the consumer's declared key list, invoked pre-provision; or a verify_plan.py-adjacent check + pytest against pinned fixture bundles)

Scope / surfaces

Primary target: scripts/verify_plan.py, scripts/ (reuse-staging helpers), tests/
Grep the workflow surface for the pattern before editing (grep -rln '<pattern>' .claude/ CLAUDE.md scripts/) and update every hit; list them in the plan.

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; if workflow.yaml or CLAUDE.md change, they stay consistent with the rule file.
This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard, workflow-fix-on-bug.md).

Provenance

workflow_fix_target: scripts/verify_plan.py, scripts/ (reuse-staging helpers), tests/
origin: parked candidate on task #1087 at 2026-07-06T11:41:55Z

Verbatim parked note:

parked — running under workflow_fix_target Provenance / EPM_WORKFLOW_FIX_SESSION (recursion guard, see workflow-fix-on-bug.md § Recursion guard). Candidate surfaced as prose by the Alternatives critic ensemble (Claude + Codex both, non-blocking): MECHANIZE the realized-keys check — a pre-provision key-schema probe (torch.load(mmap=True).keys() superset-of consumer asserts) in the reuse-staging path, or a pytest running the consumer loader against pinned bundles; target_file: scripts/ reuse-staging helpers / verify_plan.py-adjacent check; confidence: medium; related_task: #1087 (origin #1073). NOT auto-routed by this session; parked for the next human/orchestrator pass (nightly /daily sweep will surface it).