workflow-fix: gotchas bullet — single-arm smokes blind to per-arm seams

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a gotcha_candidate: yes failure-lesson raised on task #1090 fu5 (emitting agent: experiment-implementer).

Goal

add a gotchas bullet: per-arm-class smoke coverage — a driver extended to a new source-context class crashes the #527/#538 panel-disjointness assert at ModelOrganism sites a single-arm smoke never reaches; thread the source-filtered panel and smoke one run per arm class

Workflow gap

• Bug observed: #1090 fu5: all 3 imp-bare arms trained to completion then died rc=2 at the ladder entry on the panel-disjointness AssertionError; the fu5 smoke default covered only the formatting arm
• Why it is a workflow gap: the tiny-real smoke standard (gotchas "Mock-seam smokes" entry + Step 6d.0-bis) does not state that smoke coverage must span ARM CLASSES — a per-arm seam (context / negative-panel / organism assembly) is invisible to a single-arm smoke, and this cost a full 4xA100 GCE cycle. The library refusal itself is correct and load-bearing (contrastive-negatives disjointness invariant).
• Confidence (emitter): high
• verified-at-filing: grep -n "one run per arm class\|arm class\|panel_name_for" .claude/rules/gotchas.md → 0 hits for the proposed guidance in the named target (absence-of-guard claim — the 0-hit result IS the evidence); the invariant itself lives in .claude/rules/contrastive-negatives.md § Disjointness invariant (present) (2026-07-15)

Proposed change (candidate diff sketch — refine in planning)

• gotchas.md, near the "Mock-seam smokes surface production shape bugs" entry:
• • Single-arm smokes are blind to per-arm seams. A driver extended to a NEW
• source-context class (e.g. bare/"default") can pass every existing smoke and
• still crash at ModelOrganism construction: the #527/#538 panel-disjointness
• assert refuses a source content-identical to a default-panel member. Thread
• the source-filtered panel (fu3w.panel_name_for-style) at every organism site
• the new class reaches, and make the smoke default cover ONE RUN PER ARM
• CLASS, not one run total (#1090 fu5, a full GCE cycle lost).

Scope / surfaces

• Primary target: .claude/rules/gotchas.md
• Consider a one-line cross-ref from Step 6d.0-bis (tiny-real standard) in .claude/skills/issue/SKILL.md if the planner judges it in-scope.

Constraints / invariants

• Workflow-surface only; workflow_lint passes (incl. --check-lessons-index if a rules file is added — this edits an existing rule, no index change expected).
• This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — recursion guard applies.

Provenance

• workflow_fix_target: .claude/rules/gotchas.md
• fingerprint: ac8fb68b9f65

failure_class: code phase: issue1090_fu4.py --phase run (Tier-1 ladder entry, fu5 bare-context arms) lesson: A driver re-parametrized for a NEW source-context class (bare/"default") can pass every existing smoke yet crash at ModelOrganism construction, because the #527/#538 panel-disjointness invariant refuses a source that is content-identical to a default-panel member — thread the source-filtered panel (fu3w.panel_name_for) at EVERY ModelOrganism site a new context class reaches, and make the smoke default cover one run per ARM CLASS, not one run total: a per-arm seam (context/panel/organism assembly) is invisible to a single-arm smoke. generalizes: yes owning_agent: experiment-implementer gotcha_candidate: yes root_cause_confirmed: yes