`### Blocking` needs a producer contract or a model-judged emptiness rule — regex-ising an LLM-written list field repeats the #316 round-1 error

[kiki830621]

Problem

### Blocking has no producer contract. idd-update writes it from the template - {blocker 1, or "(none)"}, the {} is model-filled, and across this repo's 55 real sections the 48 semantically-empty values use 35+ spellings (- (none), - (none — closed), - (none) — diagnosed, awaiting pickup, （無）, - (none — ready for /idd-implement #107) …).

On main the emptiness of this section was read as prose by the model (idd-list Step 3.7) and judged correctly. PR #318 round 2 replaced that with a regex in idd_blocking_section — which misjudged 31 of 48 empty sections as live blockers — and at the same time promoted the signal from an idd-list display grouping to a hard abort in four skills. Round 3 of that PR stops the bleeding with a leading-token rule plus a frozen corpus-blocking.json, but the Devil's Advocate lens made the structural point: stacking character classes on an LLM-written field is the same category error round 1 made on ### Complexity, and two attempts at the regex each grew a new hole.

Source

• Verify report on PR feat: actionability gate round 2 — prefix tier + deferral vocabulary, wired into all four consumers (#316) #318 (round 2), finding 19 (DA-HIGH-2): PR feat: actionability gate round 2 — prefix tier + deferral vocabulary, wired into all four consumers (#316) #318 discussion → verify master comment.
• Related: bug: #298 的修正只落在 idd-list — 另三個 Complexity consumer 未動，且 idd-list 自身 Step 5 與 Step 3.7 互相矛盾 #316 (the gate), feature: /idd-list blocked-state awareness output mode (anti-anxiety surfacing) #84 (original blocked-state surfacing), feat: parked issue 無回訪機制 — ic-r011 宣稱的 periodic backlog grooming 不存在，且它 grep 的 blocker:* label 零使用 #310 (parked revisit).

Expected

Pick one of two shapes and record the decision in references/actionability-gate.md:

1. Producer contract: idd-update (and every writer of ### Blocking) emits a machine-checkable token — e.g. an empty section contains no bullet at all, or exactly - (none) with annotations moved to the next line — and the reader is then allowed to be strict. Needs a migration note for the 48 existing values (or a one-time normalisation pass).
2. Model judgement: the shared helper returns the section's raw lines; the four consumers keep a single shared extraction but the emptiness verdict is made by the executing model against a written rubric, with the raw lines surfaced.

Either way: the choice must be validated against the 55-row corpus-blocking.json shipped by #316, and the signal-3 row of §Risk posture must state which failure direction the chosen shape accepts.

Acceptance criteria

• Decision recorded in references/actionability-gate.md with the rejected alternative and why
• If (1): producer template updated, existing values handled, reader tightened, corpus fixture updated
• If (2): helper exposes raw lines, rubric written, four consumers updated, corpus fixture remains the regression baseline
• corpus-blocking.json 0 FP / 0 FN under the chosen shape

Complexity

Plan — cross-file (helper + 4 consumers + idd-update template + reference), decision-heavy between two valid shapes, no new published API.

[kiki830621]

Note — items handed to this issue by the PR #318 round-3 verify (2026-09-07)

Source: verify master report on PR #318 (round 3), rows 8, 14, 17, 18 and the Devil's Advocate §4 premise ruling. These were deliberately not fixed in round 4 because each one is "how should ### Blocking be read", i.e. this issue's question; round 4 only restated the accepted-miss rule and pinned both failure directions by test.

Measured facts the decision here must absorb

• The frozen corpus (fixtures/corpus-blocking.json, 55 rows, now carrying original bodies) is 54/55 CLOSED issues; the gate skips non-open issues. On the 2026-09-07 open backlog signal 3 decided exactly one issue (bug: #298 的修正只落在 idd-list — 另三個 Complexity consumer 未動，且 idd-list 自身 Step 5 與 Step 3.7 互相矛盾 #316). The corpus is a valid sample of producer style, not evidence of gate correctness on the live backlog — the acceptance bar for either shape chosen here is the semantic truth (48 empty / 7 real blockers), not the fixture's expect_empty (row idd-all: 增加 HITL（attended）模式 — direct-commit + 允許 sub-skill AskUserQuestion #1 is a rule-accepted false positive).
• The leading-token rule's edge behaviour, measured: - (none) followed by an ordered-list item / + bullet / blockquote / table row / #### line / bare paragraph reads empty (fail-open); a lead-in sentence before - (none) reads as a blocker (fail-closed); identical text is a blocker on line 1 and invisible on line 2. Widening the bullet class was measured and rejected in round 3: it enlarges the fail-closed side.
• Only the first ### Blocking section of a body is read ([enhancement] acceptance-review 缺在 IDD mode model — 「人不 review」需要 first-class 做法 #102 has four, IDD human-in-the-loop should conform to the NSQL confirmation protocol (v4.1.0) #103 two — stale-first wins). idd-update semantics say "current", i.e. arguably the last.
• IC_R011 (c) "blocked-on-external" now files with parking-lot (after blocker:* was retired) and therefore lands in the parked group, while its meaning is the gate's blocked; if the field gets a producer contract, (c) should write the external condition into the new issue's ### Blocking instead.
• Two candidate placeholder regexes were measured and rejected; their text and FP/FN counts are recorded in scripts/lib/actionability.sh so they are not re-derived.

Whichever shape is chosen — producer token or model judgement — should be validated against the 55-row corpus (through the extractor, with bodies) and against the open backlog, and should decide first-vs-last section and the (c) semantics in the same change.