EPS Dashboard

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate the orchestrator raised while driving /issue 2502 (autonomous). The Step 6d.0 PASS_AUTHORIZED_STUB mechanical grant REFUSED with "no per-arm-resolution: sub-block found" even though the marker carried a substantively-correct, line-anchored per-arm-resolution: sub-block with all 15 rows present.

Goal

Make the authorized-stub / smoke-architecture per-arm-resolution row parser tolerant of the — (em-dash) arm/resolution separator that implementers of <driver>:<phase> argparse-choices drivers naturally write, and document the backtick-colon requirement in the implementer spec — so a grammar-only malformation of correct content is caught earlier (or parses) instead of surfacing only at Step 6d.0.

Workflow gap

Bug observed: check-authorized-stub REFUSED PASS_AUTHORIZED_STUB with "no per-arm-resolution: sub-block found" because the per-arm rows used — (em-dash) as the arm/resolution separator (- gen_capture:gen — FALLBACK — <detail>) with unbackticked colon-containing arm names, and _PER_ARM_ROW_RE (src/explore_persona_space/task_workflow.py:2586) requires a : separator and — for arm names containing an internal colon — backticks; the em-dash rows parse per_arm to {}.
Why it is a workflow gap: _PER_ARM_ROW_RE accepts a colon-containing arm name ONLY via its backticked alternative (`[^`]+`), but the implementer spec's per-arm-resolution example (.claude/rules/experiment-implementer-section-reference.md:92, <arm-name>: REAL — <detail>) shows the UNBACKTICKED colon form and never warns that a <driver>:<phase> arm name needs backticks. Implementers of the argparse-choices driver class (explicitly a #2502 pattern) naturally render <driver>:<phase> — RES — detail, which silently fails to parse. The failure is invisible on the PASS_PARTIAL path (that path never parses per-arm rows) and surfaces only at the Step 6d.0 PASS_AUTHORIZED_STUB grant — the worst place to discover it. This class has 9+ documented hits in .claude/agent-memory/code-reviewer-lean/smoke_arch_marker_2176_grammar_pitfalls.md.
Confidence (emitter): medium
verified-at-filing: grep -rn '_PER_ARM_ROW_RE' src/explore_persona_space/task_workflow.py → 3 hits (def at :2586 + 2 usages at :2704 doc, :2748 call); grep -rln 'per-arm-resolution' .claude/agents/experiment-implementer.md .claude/rules/experiment-implementer-section-reference.md → both present (spec sites) (2026-08-24)

Proposed change (candidate diff sketch — refine in planning)

- _PER_ARM_ROW_RE = re.compile(r"^\s*-?\s*([^:`*]+?|`[^`]+`)\s*:\s*(REAL|FALLBACK|N/A)\b")
+ # Accept BOTH the canonical colon separator AND the em-dash separator
+ # implementers of <driver>:<phase> argparse-choices drivers naturally write.
+ # Arm-name group: a backticked span (any chars incl. ':'), OR a bare run that
+ # may contain ':' when the separator is em-dash.
+ _PER_ARM_ROW_RE = re.compile(
+     r"^\s*-?\s*(`[^`]+`|[^`*]+?)\s*(?::|—|–|-)\s*(REAL|FALLBACK|N/A)\b"
+ )

Refine in planning: the bare-name alternative must not greedily swallow the resolution token; anchor the em-dash/colon separator so REAL|FALLBACK|N/A is the token immediately after it. Add a fixture reproducing the - gen_capture:gen — FALLBACK — detail shape (parses to arm gen_capture:gen) and keep the existing colon + backticked cases green. Mirror the clarification into .claude/agents/experiment-implementer.md + .claude/rules/experiment-implementer-section-reference.md (colon-containing arm names: either backtick under the colon form, or use the em-dash form).

Scope / surfaces

Primary target: src/explore_persona_space/task_workflow.py (_PER_ARM_ROW_RE, parse_smoke_arch_marker, authorized_stub_grant)
Spec docs: .claude/agents/experiment-implementer.md, .claude/rules/experiment-implementer-section-reference.md
Grep the surface for the pattern before editing and update every hit; list them in the plan.

Constraints / invariants

Workflow-surface only — never experiment code, configs/, or tasks/.
Existing colon-form and backticked-form per-arm rows MUST keep parsing (regression fixtures for both).
scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; tests/test_workflow_fix_dedup.py + any test_task_workflow* stay green; if the spec docs change they stay consistent with the parser.
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).

Provenance

workflow_fix_target: src/explore_persona_space/task_workflow.py,.claude/agents/experiment-implementer.md,.claude/rules/experiment-implementer-section-reference.md
fingerprint: 8643a11b43e7

target_file: src/explore_persona_space/task_workflow.py, .claude/agents/experiment-implementer.md, .claude/rules/experiment-implementer-section-reference.md bug_observed: Step 6d.0 PASS_AUTHORIZED_STUB grant REFUSED "no per-arm-resolution: sub-block found" because per-arm rows used the em-dash arm/resolution separator with unbackticked colon-containing arm names, which _PER_ARM_ROW_RE parses to empty. why_workflow_gap: _PER_ARM_ROW_RE accepts colon-containing arm names only when backticked, but the implementer spec's per-arm-resolution example shows the unbackticked colon form and never warns that a : arm name needs backticks; implementers of argparse-choices drivers naturally write the em-dash form, which silently fails to parse and surfaces only at the Step 6d.0 grant. proposed_change: Make _PER_ARM_ROW_RE accept the em-dash arm/resolution separator (- — RES — detail) in addition to the colon form, so colon-containing : arm names parse without backticks; document the backtick-colon requirement in the implementer spec. diff_sketch: |

_PER_ARM_ROW_RE = re.compile(r"^\s*-?\s*([^:*]+?|[^]+)\s*:\s*(REAL|FALLBACK|N/A)\b")

_PER_ARM_ROW_RE = re.compile(r"^\s*-?\s*([^]+|[^]+?)\s(?::|—|–|-)\s*(REAL|FALLBACK|N/A)\b") confidence: medium related_task: #2502