Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a <!-- workflow-fix-candidate v1 --> block raised on task #1774 (emitting agent: clean-result-critic, round 1).

Goal

Widen the H-code token regex in verify_task_body.py check 28 to match H? hypothesis-tag forms (e.g. \bH\d+[a-z]?\b) in figure sidecar text.

Workflow gap

• Bug observed: The figure-text opaque-config-code check (check 28) PASSed task #1774's body although the jensen figure sidecar registered title_left: "Jensen-gap direction concentration (H1c)" — the H-code pattern \bH\d\b misses the H form (\b cannot fire between the digit and the trailing letter, so H1c never matches).
• Why it is a workflow gap: The mechanical backstop for plan-internal hypothesis codes in figure text exists (check 28, _HYPOTHESIS_CODE_RE) but its token pattern is narrower than the tags plans actually emit (H1a/H1c/H4b...), so the exact class it was built to catch passes silently; the miss cost a clean-result-critic round on #1774 (r1 blocker 2: figure regenerated + body re-pinned).
• Confidence (emitter): medium
• verified-at-filing: grep -n 'H\\d' scripts/verify_task_body.py → _HYPOTHESIS_CODE_RE = re.compile(r"\bH\d\b") at scripts/verify_task_body.py:7816, consumed by _opaque_code_tokens (:7834) in check 28 (:7794/:7964); semantic probe: re.search(r"\bH\d\b", "concentration (H1c)") → None (the H1c form passes the current regex), 1 target file, presence hit read in context — the hit IS the too-narrow pattern, not a landed fix (2026-07-29)

Proposed change (candidate diff sketch — refine in planning)

- _HYPOTHESIS_CODE_RE = re.compile(r"\bH\d\b")
+ _HYPOTHESIS_CODE_RE = re.compile(r"\bH\d+[a-z]?\b")
+ # apply to sidecar title/axis/legend text as today; keep the existing
+ # standalone/parenthesized-token guard to bound false positives

Scope / surfaces

• Primary target: scripts/verify_task_body.py
• Grep the workflow surface for the pattern before editing (grep -rn '_HYPOTHESIS_CODE_RE\|\\bH..d..b' scripts/ .claude/) and update every hit; list them in the plan. Add a regression test (H1c-form token in a fixture sidecar must FAIL check 28; a legitimate "H100"-style hardware token must not false-fire — calibrate the [a-z]? bound + surrounding-token guard accordingly).

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).

Provenance

• workflow_fix_target: scripts/verify_task_body.py
• fingerprint: 531b88a3a471

target_file: scripts/verify_task_body.py bug_observed: The figure-text opaque-config-code check (advertised classes "slug / @L-pin / H-code / slot-family tokens") PASSed task #1774's body although the jensen figure sidecar registers "title_left":"Jensen-gap direction concentration (H1c)" — the H-code pattern misses the H<digit><letter> hypothesis-tag form. why_workflow_gap: The mechanical backstop for plan-internal hypothesis codes in figure text exists but its token pattern is narrower than the tags plans actually emit (H1a/H1c/H4b...), so the class it was built to catch passes silently. proposed_change: Widen the H-code token regex in the figure-text opaque-code check to match parenthesized/standalone \bH\d+[a-z]?\b in sidecar title/axis/legend text. diff_sketch: |

• H_CODE_RE = re.compile(r"\bH\d\b") # (current form misses H1c)

• H_CODE_RE = re.compile(r"\bH\d+[a-z]?\b")

• apply to sidecar .text.axes[*].title_left / suptitle as well as legend/axis labels,

• keeping the existing standalone/parenthesized-token guard to bound false positives

confidence: medium related_task: #1774