workflow-fix: gotchas.md — companion-stat drop-class semantics (zero split-half noise floors)

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a gotcha_candidate: yes failure-lesson raised on task #1415 (emitting agent: experiment-implementer via the crash-fix failure-lesson block; router: orchestrator).

Goal

Add a gotchas.md bullet: companion statistics need their own drop-class semantics; zero split-half noise floors at shared-first-token early positions are a named non-fatal exclusion, never an integrity-guard trip.

Workflow gap

• Bug observed: the issue1415 p3 driver died rc=1 because a >20% drop guard calibrated for the registered statistic false-fired on the Delta_floor companion's legitimate zero-floor degeneracies.
• Why it is a workflow gap: .claude/rules/gotchas.md (the on-demand trap catalog loaded when writing analysis/eval code) has no entry for this trap class; the agent-memory copy reaches only experiment-implementer spawns, not the wider gotchas load surface.
• Confidence (emitter): high (root_cause_confirmed: yes in the epm:failure-lesson v1 block on #1415)
• verified-at-filing: grep -c "noise floor\|nonpositive\|drop-class\|companion stat" .claude/rules/gotchas.md → 0 hits (2026-07-22) — absence-of-guard claim, 0-hit in-target IS the evidence.

Proposed change (candidate diff sketch — refine in planning)

• gotchas.md new bullet (analysis/eval section): split-half noise floors of
• teacher-forced activations are EXACTLY zero at early answer positions when all
• draws share their first tokens (deterministic openings — 7/28 pairs on #1415);
• log-ratio/guard code over floor magnitudes must treat non-positive means as a
• NAMED non-fatal exclusion class; companion statistics get their OWN drop-class
• semantics + thresholds (a fail-loud guard calibrated for the registered
• statistic false-fires on a companion's legitimate degeneracies); regression-test
• with a shared-first-token fixture.

Scope / surfaces

• Primary target: .claude/rules/gotchas.md
• Cross-check the LESSONS.md index row for gotchas.md stays accurate (--check-lessons-index).

Constraints / invariants

• Workflow-surface only. scripts/workflow_lint.py no-flags run passes (lessons-index + size budgets).
• 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: f57b29a31bb0

target_file: .claude/rules/gotchas.md bug_observed: the issue1415 p3 driver died rc=1 because a >20% drop guard calibrated for the registered statistic false-fired on the Delta_floor companion's legitimate zero-floor degeneracies why_workflow_gap: gotchas.md has no entry for the zero-noise-floor / companion-stat drop-class trap, so future analysis code will re-hit it outside experiment-implementer's agent memory proposed_change: add a gotchas.md bullet: companion statistics need their own drop-class semantics; zero split-half noise floors at shared-first-token early positions are a named non-fatal exclusion, never an integrity-guard trip diff_sketch: |

• new gotchas.md bullet per the epm:failure-lesson v1 block on #1415 (see Goal) confidence: high related_task: #1415