EPS Dashboard

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a gotcha_candidate: yes failure-lesson raised on task #1315 (emitting agent: experiment-implementer, crash-fix round 7).

Goal

Extend the gotchas.md teacher-forced-capture BPE-seams entry: under prefix_end='last_user' a plain-text span boundary with a space-before-{q} wrap BPE-merges on essentially every question (not an edge case), and span-rig smokes must include at least one plain-text-boundary context (special-token-adjacent-only slices cannot catch the class).

Workflow gap

Bug observed: #1315 r7: neg_reph_curious ("I'm curious about the following: {q}") drifted 20/20 rows at the prefix boundary under prefix_end='last_user' while all 160 special-token-adjacent rows were exact; the smoke slice contained no plain-text-boundary context so the class first fired in production p8_capture.
Why it is a workflow gap: gotchas.md carries the BPE-seams entry (the #1092 offset-mapping recipe) but is silent on (a) plain-text (non-special-token-adjacent) boundaries merging on ~every question rather than rarely, and (b) the smoke-composition duty — a span-rig smoke must include a plain-text-boundary context or the class is invisible until production.
Confidence (emitter): high (root_cause_confirmed: yes; fix + 20/20-vs-160/160 CPU repro landed on issue-1315 @ 7075942917)
verified-at-filing: grep -ciE "BPE seam|bpe.seam|offset.mapping" .claude/rules/gotchas.md → 1 hit (presence of the existing entry confirmed); grep -ciE "plain.text.boundar|last_user|space before" .claude/rules/gotchas.md → 0 hits (absence-of-nuance claim — the 0-hit result IS the evidence) (2026-07-15)

Proposed change (candidate diff sketch — refine in planning)

In .claude/rules/gotchas.md, extend the existing teacher-forced capture / BPE-seams entry:

Plain-text boundaries are the WORST case, not an edge case: under
prefix_end='last_user' any "... {q}"-style wrap with a space before the
query BPE-merges the boundary into the query's first word on essentially
every question (#1315 r7: 20/20 rows for one panel member vs 160/160 exact
special-token-adjacent rows). Derive both boundaries from the full render's
offset mapping (exclude a prefix-boundary straddler, include a
context-boundary straddler) with per-row seam provenance; and a span-rig
SMOKE must include at least one plain-text-boundary context — a
special-token-adjacent-only slice cannot catch the class.

Scope / surfaces

Primary target: .claude/rules/gotchas.md
Grep before editing (grep -rln 'BPE' .claude/rules/ .claude/agents/) and keep consistent with the existing #1092 recipe text; list hits in the plan.

Constraints / invariants

Workflow-surface only. scripts/workflow_lint.py no-flags run passes.
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: 21c9cb462625

failure_class: code phase: p8_capture (compute_prompt_spans span validation) lesson: Under prefix_end='last_user' a span boundary can sit on PLAIN TEXT inside the user turn, and any "... {q}"-style wrap with a space before the query BPE-merges that space into the query's first word on essentially every question — re-tokenizing text[:boundary] then asserting token-prefix identity is guaranteed to fail there. Derive both boundaries from the FULL render's offset mapping (exclude a prefix-boundary straddler, include a context-boundary straddler) with per-row seam provenance; keep the full-sequence token-identity assert for genuine render/tokenizer drift. Smokes for span rigs must include at least one plain-text-boundary context (a special-token-adjacent-only slice cannot catch this class). generalizes: yes owning_agent: experiment-implementer gotcha_candidate: yes root_cause_confirmed: yes