workflow-fix: smoke must cover every behavior-class x regime (smoke/parity family)

Overview / Motivation

Auto-filed from a deep-dive on #1586's crash history (2026-07-23, user-directed). #1586's pre-launch smoke ran ONE content-class cell (syc-pers-ft-con-s137), so the marker-class, positive-only (po), and panel read-side code paths were NEVER exercised before the full run. Three distinct bug classes then surfaced LIVE, one per phase, over hours (whack-a-mole): r3 read-side organism panel disjointness, r4 marker po mix-rowcount (200 vs copy-filled 1000), r6 reuse-seam schema. A smoke that covered every (behavior-class x regime) would have caught all three pre-launch.

Goal

Add a "regime/class-coverage" member to the smoke/production-parity family in gotchas.md so multi-regime subprocess-per-phase dispatchers exercise every class-specific code path in the smoke, not just one content-class cell.

Workflow gap

• Bug observed: issue1586 pre-launch smoke ran one content-class cell so marker-class, positive-only, and panel read-side paths were unexercised and three bug classes (r3 read-side panel disjointness, r4 marker po mix rowcount, r6 reuse-seam) surfaced live one per phase.
• Why it is a workflow gap: the smoke/production-parity family in gotchas.md covers gate calibration (#1345) and process width (#1315/#1333) but has no member requiring COVERAGE across (behavior-class x regime); a dispatcher spanning multiple behavior classes/regimes must smoke >=1 cell per class x regime.
• Confidence (emitter): high
• verified-at-filing: grep -niE "full-panel|fresh-child smoke|subprocess-per-phase|regime|behavior-class" .claude/rules/gotchas.md (2026-07-23) → the smoke/production-parity family exists (gate-calibration #1345, process-width #1315/#1333) but NO class×regime-coverage member.

Proposed change (candidate diff sketch — refine in planning)

Add to the smoke/production-parity family in .claude/rules/gotchas.md:

A subprocess-per-phase dispatcher spanning MULTIPLE behavior classes (e.g. marker
vs content) and/or regimes (e.g. contrastive-negatives `con` vs positive-only `po`)
MUST smoke >=1 cell per (class x regime), so each class's distinct code paths run
pre-launch (marker parity read, `po` mix asserts, panel disjointness read) — not
just one content-class cell. Incident #1586 r3/r4/r6.

Scope / surfaces

• Primary target: .claude/rules/gotchas.md (smoke/production-parity family).
• Consider a one-line cross-ref from .claude/skills/issue/SKILL.md Step 6d.0 (smoke) if the planner judges it warranted.

Constraints / invariants

• Workflow-surface doc only; no experiment code.
• scripts/workflow_lint.py passes; if the lessons index or a sibling rule is touched, they stay consistent.
• Runs under EPM_WORKFLOW_FIX_SESSION=1 (recursion guard).

Provenance

• workflow_fix_target: .claude/rules/gotchas.md
• fingerprint: b5fabf456761

Origin: user chat 2026-07-23 ("Fix all the problems in the background with happycoder") on the #1586 crash-history review. Related task: #1586 (r3/r4/r6).