workflow-fix: gotchas entry — BPE-seam position-parity gate tail (carve-out, not threshold-loosen)

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a gotcha_candidate: yes failure-lesson raised on task #825 (emitting agent: experiment-implementer, round-10 crash-fix).

Goal

Add a gotchas.md entry: BPE delimiter-seam merges under plain-text "Role: content" renders shift span-derived positions ±1 between two captures, breaking cross-capture position/activation-parity gates on a mechanically-identifiable row tail (median clean, row-fraction leg fails); chat-template renders are immune.

Workflow gap

• Bug observed: #825 round-10's G2 activation-parity gate HALTed on pretrained (frac 0.9767 < 0.99) — 60/2572 rows where context_pos shifted ±1 vs the banked store at the "Assistant: " BPE seam; cosines 0.687–0.872 on exactly those rows.
• Why it is a workflow gap: gotchas.md carries the zero-width-span BPE-seam trap but not its position-PARITY-gate sibling; the next cross-capture parity gate over a plain-text render will re-hit it and a naive fix loosens the threshold (selection-on-outcome risk).
• Confidence (emitter): high

Proposed change (candidate diff sketch — refine in planning)

• gotchas.md § BPE zero-width spans entry gains a sibling paragraph: cross-capture
• position-parity gates (G2-style) fail on the seam-shifted row tail; diagnose by
• comparing position METADATA across captures (new pos vs banked token_start) BEFORE
• touching thresholds; resolve with a mechanical position-keyed pair-safe carve-out
• (fail-loud rate cap), never a cosine-keyed exclusion or fraction loosening.
• Signature: instruct(chat-template)-clean / pretrained(plain-text)-tail asymmetry.
• Worked example: scripts/issue825_onpolicy_turn_depth_fit.py::_boundary_pos_carveout.

Scope / surfaces

• Primary target: .claude/rules/gotchas.md

Constraints / invariants

• Workflow-surface only; workflow_lint passes.

Provenance

• workflow_fix_target: .claude/rules/gotchas.md
• fingerprint: 6c22949edabb