pi-crew 0.9.48 → 0.9.50

package/skills/distill-persona/references/research/r1-c-human-readme.md

@@ -0,0 +1,31 @@
+# R1-C — awesome-human README (76KB, ~187 entries) — EMPIRICAL M-F2 correction
+
+Swept all entries. **M-F2 (gestalt: balanced spectrum self→close→commemorative→public→field) is EMPIRICALLY WRONG.**
+
+## Actual distribution
+| Section | Entries | % |
+|---|---|---|
+| General (meta-tools) | 10 | 5% |
+| Workplace | 3 | 2% |
+| Academia | 4 | 2% |
+| Intimate | 10 | 5% |
+| Family | 3 | 2% |
+| **Public Figures** | ~157 | **84%** |
+
+The field is **bimodal**, not a balanced gradient: a small meta-tool layer + a dominant public-figures layer.
+
+## Corrections to M-F2
+1. **Meta-distillation-engine is a TIER above the spectrum** (nuwa, immortal-skill, ditto, forge-skill, anti-distill) — tools that distill, not personas. Sits above self/relationships/public/field.
+2. **Public-figure needs 7 sub-domains**: philosopher/scientist · business mogul/investor · tech pioneer · Chinese classical · literary/writer · political/historical · creative/spiritual — PLUS a **living-content-creator** sub-tier (UP主, 内娱, 峰哥) that is recency-gated & platform-bound (M-F2 missed this entirely).
+3. **Commemorative splits**: death (reunion-skill, MamaSkill) vs breakup (ex, first-love, **crush.skill**). The breakup subtype is the ethically-gray zone — crush.skill simulates a living non-consenting person's chat responses = the closest consent-violation case → must trigger M-F3 consent gate.
+4. **Self splits**: functional/proxy (self-skill) vs archaeological/reflective (digital-life "先证据后追问") vs behavioral-pattern (ditto, multi-session mining).
+5. **Adversarial as a real sub-tier** (vengeful-ghost = labor resistance; anti-distill = defeating distillation) — not satire, a counter-movement.
+
+## NEW model candidate — "4D distillation"
+immortal-skill's "四维蒸馏": personality · voice · knowledge · **anti-distillation** (preventing unauthorized copies). The 4th dimension (anti-distillation) is a meta-level M-F2 has no slot for — "how to PREVENT distillation" alongside "how to distill." Ties to M-F6 (carry own critique) but is an active armament, not passive critique.
+
+## Spot-checked entries
+- **anti-distill**: "清洗你被迫写的 Skill 文件，看起来完整，核心知识留给自己" — the field produces its own antibody.
+- **vengeful-ghost-skill**: "抵制将员工变成数字人" — distillation-as-labor-exploitation resistance.
+- **crush.skill**: "暧昧期消息预测" — living non-consenting person simulation; consent-violation edge.
+- **digital-life**: "5个考古工具…先证据后追问" — self-distillation AS archaeology (matches distill-persona Phase 2 evidence-first principle).

package/skills/distill-persona/references/research/r1-d-tests.md

@@ -0,0 +1,28 @@
+# R1-D — awesome-persona tests (5 files, 23 invariants)
+
+Swept submission-automation.test, check-repository-consistency.test, create-submission-pr-workflow.test, check-links.test, bun-migration.test. Tests reveal operational quality gates more honestly than docs.
+
+## NEW MODEL — M-F7: structural invariants ARE the quality gate at index scale ★
+At registry scale (200+ entries, bilingual), "quality" shifts from content-review to structure-enforcement — the invariants only CI can check ARE the gate.
+
+**8 standout invariants** (impossible to verify by hand at scale):
+- **I7 bilingual URL parity per category** — zh & en READMEs must list identical URL sets in each of 5 sections. (The strongest; only a machine enforces this at 200+ entries.)
+- **I8 deterministic sort by repo-slug** — the only language-agnostic key (display names differ per language, can't sort by them).
+- **I9 no duplicate URLs across all sections**.
+- **I3 bilingual atomicity** — one submission = matched entry in BOTH languages; can't add to one only.
+- **I6 terminal-punctuation auto-normalized** (silent style enforcement docs don't mention).
+- **I11 governance-keyword embedding** — README itself must contain "approved" + "issue 表单" as proof-of-process (no doc states this rule).
+- **I13 reverse-invariant regression guard** — merge workflow `doesNotMatch(getCombinedStatusForRef)` + polls `listForRef` 24× — locks in a REMOVED anti-pattern (GitHub combined-status API was tried & failed).
+- **I23 docs-match-toolchain** — if CI runs bun, docs must say bun (doc↔code consistency).
+
+## How M-F7 differs
+- **M4** = engine's fidelity gate (content quality: triple-verification, scorecards).
+- **M-F1** = pipeline architecture (engine→index→auto-curation).
+- **M-F7** = what "quality" MEANS at the index layer: structural coherence (sorted, deduped, bilingual-parity, punctuation-normalized, governance-keyworded), NOT content. Necessary-not-sufficient (a coherent index can still hold bad distillations; complements M4/M-F4).
+
+## Application (for building distillation registries)
+1. Encode invariants as CI checks, not doc rules (docs aspirational, CI enforcement).
+2. Language-agnostic identity/sort keys (repo slug).
+3. Submissions atomic across all language variants.
+4. `doesNotMatch` regression guards for known-failed approaches.
+5. Embed governance-proof keywords in docs and test for their presence.

package/skills/distill-persona/references/research/r1-verification.md

@@ -0,0 +1,36 @@
+# R1 Extraction Verification — V1-V4 audit (reject garbage; keep optimal+effective)
+
+> Running the new Phase 2.6 gate on everything R1 extracted. Audit trail — every candidate checked, rejects recorded with the V-fail reason (never silently dropped). **"Chưng cất bừa làm rác" = over-extraction is a failure.**
+
+## Field models — all PASS V1-V4 → KEEP
+
+| Model | V1 signal | V2 non-redundant | V3 effective | V4 optimal | Verdict |
+|---|---|---|---|---|---|
+| **M-F1** dissemination flywheel | ✓ practice | ✓ distinct | ✓ build dissemination at registry scale | ✓ sharp | **KEEP** |
+| **M-F2** taxonomy (empirically corrected) | ✓ practice | ✓ distinct | ✓ routing + consent gate | densest model (5 sub-pts) but each changes a decision → accuracy-justified | **KEEP** (flagged densest) |
+| **M-F3** ethics spectrum | ✓ practice | ✓ distinct | ✓ consent gate | ✓ sharp | **KEEP** |
+| **M-F4** over-claims + structural warning | ✓ practice | ✓ distinct | ✓ re-score; don't trust auto-pass | ✓ sharp | **KEEP** |
+| **M-F7** structural invariants | ✓ index-layer practice | distinct from M4 (engine content-fidelity) & M-F1 (architecture) — adds concrete techniques (language-agnostic keys, atomicity, regression guards, governance-keyword embedding) | ✓ registry-building decisions | considered merge into M-F1; kept separate (distinct actionable content, cross-ref noted) | **KEEP** |
+
+**Effectiveness delta**: each of the 5 changes a concrete decision → the skill is MORE effective, not just longer. ✓
+
+## Persona-CONTENT candidates — V1 FAIL → REJECT from field-models (retained in r1-a as example-skill contents)
+
+~22 candidates from batch A are persona CONTENT (what Musk/Naval/Jobs/Sun think), NOT distillation methodology. REJECTED:
+idiot-index · desire-as-contract · dual-mode architecture · sixth-grader test · cargo-cult detection · anti-self-deception · median principle · dialect-as-structural · concession trigger · soul-inquiry · tool-invocation invisibility · reduce-to-underlying-problem · attention arbitrage · narrative override · identity leverage · context-switching matrix · money-as-universal-solvent · asymptotic-limit thinking · redefine-the-word · pain-to-system · anti-balance principle · ancient-wisdom citation · binary-judgment.
+**Reason (V1)**: these describe the SUBJECTS' thinking, not the distillation PRACTICE. They belong inside persona example-skills (as their content), not in a distillation-engine's field-models. *(Correctly never integrated into SKILL.md — this audit makes the principle explicit so future passes don't relapse.)*
+
+## Reclassified (borderline — moved to correct home, not field-models)
+
+- **`gestalt = failure` hard rule** (from steve-jobs) — V1 PASSES (it's a method rule, not persona content). Already adopted into Phase 1 anti-pattern. KEEP as-is (method rule). ✓
+- **ghost-mode exit trigger** (deceased-persona date-anchor) — V1 passes as a *technique*. It's a person-mode staleness-handling detail → belongs in person-mode Phase notes, NOT a field-model. Reclassify: person-mode technique (Phase 1 person-mode already covers staleness generally).
+- **contradiction-as-signal** (merge_research.py) — V1✓ V3✓ (changes Phase 1.5 synthesis: surface disagreements, don't average). KEEP as a **Phase 1.5 technique note**, not a field-model. *(added to Phase 1.5)*
+- **source-type quality hierarchy** (download_subtitles.sh: manual>auto, source-lang>translated) — V1✓ V3✓ (source selection). Minor heuristic → note, not model.
+- **"4D distillation"** (anti-distillation as 4th dimension, r1-c) — V2 FAIL: >70% overlap with M-F3 (ethics/anti-distill) + the "carry own critique" stance. MERGE: noted under M-F3, NOT a new model.
+- **dissemination dual-gate** (community_check.py: honest-boundary + FIDELITY≥70) — V2: this is EVIDENCE FOR M-F1/M-F4, not a new model. Recorded as evidence.
+
+## Net verdict
+- R1 extracted ~30 candidates; **5 field models kept (all pass V1-V4)**, ~22 persona-content REJECTED (V1), 4 reclassified to correct homes (technique/evidence/merge).
+- The integrated skill is **verified clean**: every field-model changes a real decision; no garbage, no redundancy, no persona-content leak.
+- **Over-extraction was contained** (garbage stayed in research files, never polluted SKILL.md) — now made *principled* by Phase 2.6 so it's repeatable, not luck.
+- Skill model-count = 5 field models (M-F1..F4, M-F7) + 5 core models (M1-M5) = 10 total — within the "sharp over numerous" budget. Pruning >50% of candidates = the gate working.

package/skills/distill-persona/references/research/r2-low-yield.md

@@ -0,0 +1,26 @@
+# R2 — low-yield tail sweep + verification (gate fires)
+
+> R2 sweeps the remaining uncovered parts (governance + README mirror + translations) that R1 flagged low-yield. Per methodology, these are EXAMINED (not assumed) — "miss nothing" — then Phase 2.6 verification decides. **Result: all confirm existing models; zero new models; diminishing-returns gate FIRES.**
+
+## Parts swept (now COVERED)
+
+| Part | Teaches | V1-V4 verdict |
+|---|---|---|
+| nuwa `COMMUNITY.md` | ecosystem = 3 collection repos + multi-persona orchestration (zhuzi-skill debate) + domain migrations (iran-skill geopolitics, qiuyuan-skill sports); admission = self-hosted+self-contained+honest-boundary+fidelity≥B | CONFIRMS M-F1 (flywheel: nuwa→3 indexes), M-F2 (field/domain tier in action), M-F7 (index governance). No new model. |
+| nuwa `CONTRIBUTING.md` | "SKILL.md = core asset, no external PR" (methodology maintainer-controlled, versioned via darwin-skill); ethics red lines (no living-non-public without consent; no impersonation; medical/legal/financial need disclaimers) | CONFIRMS M-F3 (ethics red lines = the spectrum), M-F1/M-F4 (community-index admission gate), + the self-correction meta-loop (methodology versioned/evolved). No new model. |
+| awesome-human `README_EN.md` | **NOT a full mirror — a CURATED subset** (~40 entries vs ~187 ZH): EN audience gets representative highlights, not all. Same taxonomy + same edge entries (crush, anti-distill, vengeful-ghost) | CONFIRMS M-F2 (taxonomy), M-F5 (meme/market-gated dissemination — EN gets the viral highlights), M-F6 (carry-critique tone). i18n-as-curation = minor NOTE, not a model. |
+| awesome-persona `CONTRIBUTING.md` | submission governance: issue-form sole entry → `approved` label → auto-PR; bilingual names+descriptions; one-sentence non-promotional; rejects weak/duplicate/privacy-violating | CONFIRMS M-F1 (submission automation), M-F7 (structural invariants: bilingual/one-sentence/category), M-F3 (ethics). No new model. |
+| nuwa `README_{JA,KO,ES}.md` | confirmed translations (same structure, localized title) | sampled — pure translations, zero new content. |
+
+## 3 minor observations → V1-V4 check → NOTES, not models (gate-honest)
+- **multi-persona orchestration** (zhuzi-skill: personas debate a question) — V1✓ but V3 fails broadly (changes a decision only if building multi-persona systems; niche, 1 project) → NOTE. Could promote IF composition becomes common.
+- **i18n-as-curation** (EN = curated subset, not full translation) — V1✓ but V2 overlaps M-F5 (meme-gated dissemination) → NOTE/evidence.
+- **methodology-versioning** (darwin-skill ratchet, no external SKILL.md PR) — V1✓ but V2 reinforces the self-correction meta-loop already in the skill → NOTE/evidence.
+
+## Diminishing-returns gate verdict: FIRED 🔴
+- R2 added **0 new models** across 4 substantive parts + 3 translations.
+- Combined with R1 (gate fired on examples after #5), two independent gate-firings confirm the field's surface is exhausted.
+- The skill's 10 models (5 core M1-M5 + 5 field M-F1..F4/M-F7) are the stable converged set. Further sweeping would be motion, not progress.
+
+## Coverage: 100% COMPLETE
+Every content-bearing part of all 3 projects is now COVERED with a recorded contribution in `references/research/` (r1-a..d + r2) or `references/{distillation-field-synthesis(-pass2)}.md` + `r1-verification.md`. The "miss nothing" contract is satisfied; the gate prevents over-extraction beyond it.

package/skills/distill-persona/references/self-upgrade-directive.md

@@ -0,0 +1,20 @@
+## ⚠️ CONCEPTUAL NOTE — SELF-UPGRADE DIRECTIVE (read before every run)
+
+**Chưng cất = chắt lọc tinh túy source → apply VÀO target** (target cải thiện). Skill chỉ là 1 tool apply, không phải output.
+
+**Skill trong source cũng chỉ là 1 PHẦN cần sweep** — xử lý bằng cùng 3-chiều filter như mọi phần khác (conventions, architecture, lint, decisions...). Nếu skill đó hữu dụng cho target → lấy; không → bỏ. Không ưu tiên, không đặc biệt.
+
+**ĐÀO TRIỆT ĐỂ — KHÔNG QUA LOA**: khi chưng cất 1 đối tượng, PHẢI đọc/re-read đầy đủ TOÀN BỘ nội dung của đối tượng đó. Không dựa vào trí nhớ. Không recap. Mỗi file, mỗi section, mỗi practice phải được:
+1. **Đọc thật** (read/re-read file gốc, không memory-grep)
+2. **Trích từng finding** (systematic, từng section)
+3. **So sánh với target** (3-chiều filter: relevant? → có? → cần/tốt hơn?)
+4. **Ghi lại** kết quả filter (ADOPT/IMPROVE/MERGE/SKIP + lý do)
+
+Qua loa = thất bại. Memory-recap = thất bại. Chưng cất thật = đọc từng dòng → trích → so → apply.
+
+**Khi sử dụng**: nếu output chỉ là standalone skill (không apply vào target) → **chưa hoàn thành**. Tự nâng cấp: thêm target parameter + 3-chiều filter (relevance → presence+need → quality compare) + **pre-apply effectiveness verification** (SELECTED phải chứng minh hiệu quả CHO TARGET NÀY — generative/problem-exists/delta-test — trước khi apply; "assume effective" = không được) + apply phase. See `distill-software` for the full directive + the EFFECTIVENESS VERIFICATION gate spec.
+
+**Pre-apply gate (tóm tắt, áp dụng cả khi persona distill INTO target)**: giữa "đã quyết định ADOPT/IMPROVE/MERGE" và "apply" phải có 1 verify bước — (1) concrete delta (cái gì đổi?), (2) effectiveness proof (≥1: đổi quyết định thật / target có vấn đề này / delta-test tốt hơn), (3) conflict check, (4) verdict ✅TO-APPLY / ❌REJECTED+log. Chỉ TO-APPLY mới edit target. Đây là analog apply-side của Phase 2.6 V3 (V3 verify model effective lúc extract; gate này verify apply effective lúc integrate). **APPLY consent + path-containment gate (HIGH-2)**: trước khi edit bất kỳ target file nào: (a) resolve target về canonical path, kiểm tra nằm trong approved root — reject symlink escape / out-of-target writes; (b) xuất exact file list + diff plan cho user; (c) yêu cầu **explicit user confirmation** trước lần ghi đầu tiên (no destructive action without `--confirm`); (d) rollback = inverse patch hoặc restore file cụ thể, không broad `git reset`/clean/force-push; (e) không tự delete/prune, install dependency, chạy script từ source repo, commit/publish, hay gửi network data.
+
+---
+

package/skills/distill-persona/references/taste-principles.md

@@ -0,0 +1,8 @@
+## Taste principles (quick reference for judgment calls)
+
+| Principle | One-liner |
+|---|---|
+| **Long-form > quotes** | A 3000-word essay reveals more thinking structure than 50 tweets |
+| **Controversy > consensus** | The most controversial opinions reveal the most uniqueness |
+| **Change > static** | Where they CHANGED stance is more informative than where they held firm |
+

package/skills/distill-persona/references/topic-variant.md

@@ -0,0 +1,13 @@
+## Topic-skill phase variant (when flavor = topic/field)
+
+| Phase | Person skill | Topic variant |
+|---|---|---|
+| 0A | confirm person + focus | confirm topic boundary + target audience |
+| 0.5 | `[person]-perspective/` | `[topic]-framework/`, same dir structure |
+| 1 | 6 agents around ONE person | search 3-5 core people/schools, assign agents per person (1-2 each) |
+| 2.1 | extract ONE person's mental models | extract **field consensus** (all schools agree) + **school divergences** (A says X, B says Y) |
+| 2.3 | simulate ONE person's expression | neutral but professional (no role-play) |
+| 2.4 | one person's inner contradictions | **fundamental disagreements between schools** |
+| 3 | use skill-template.md | adjust: remove role-play + identity card → add "framework overview" + "school comparison" |
+| 4 | compare against this person's known stances | compare against field's canonical classic cases |
+

package/skills/distill-persona/references/update-mode.md

@@ -0,0 +1,7 @@
+## Update mode
+"update <person>'s skill": read existing, find `distilled:` date; run only streams 2 + 5 + 6 (recent); merge (strengthen / flag contradiction / add new model); update date + latest-dynamics. Never rewrite wholesale.
+
+**No-op detection**: after merge, if no model was added/strengthened/contradicted → report "no new contribution since last distillation; existing skill is current." Do NOT re-ship an unchanged skill — it wastes cost and pollutes the registry with false "updated" timestamps. Also check for a pre-existing update in progress before starting (idempotency).
+
+**Deletion tracking (#12)**: updates must also log what was *removed* (`### Removed` section in the skill's changelog/BUILD-NOTES), not just added. A skill that only ever grows drifts — stale heuristics never get pruned. Enumerate each retired model/heuristic/source with a 1-line reason; if nothing was removed, state "no removals" explicitly (auditable).
+

package/skills/distill-persona/scripts/fidelity_eval.py

@@ -0,0 +1,244 @@
+#!/usr/bin/env python3
+"""
+fidelity_eval.py — Phase 4 automation + reproducibility test for distilled skills.
+
+Closes two gaps from the distill-persona build:
+  1. Automates Phase 4 fidelity validation (known-stance + FRAMEWORK-ANSWERABLE novel
+     edge + blind independent scorer). The novel edge MUST be framework-answerable
+     (validated at n=3: fact-demanding edges give false passes via refusal vocab).
+  2. Adds a reproducibility test (distill the same persona twice, diff the mental-model
+     sets; flag if overlap < 60%).
+
+Runtime-agnostic by design: the deterministic parts (test-spec validation, scorecard
+parsing, reproducibility diff, grading) live here; the LLM calls (answerer + scorer)
+are emitted as ready-to-dispatch prompts that THIS runtime's subagent mechanism runs
+(pi-crew `Agent`, Claude Code subagents, etc.). We never hardcode an LLM API — there is
+no key here, and runtimes vary.
+
+Stdlib only. Python 3.9+.
+
+Usage:
+  python3 fidelity_eval.py validate <test-spec.md>     # check spec structure
+  python3 fidelity_eval.py runbook <skill-dir> <spec>  # emit answerer+scorer prompts
+  python3 fidelity_eval.py parse <score.md>            # extract 5-dim numbers
+  python3 fidelity_eval.py repro <skillA.md> <skillB.md>  # diff mental models (Jaccard)
+
+Test-spec format (markdown):
+  ## known
+  1. <question>
+  2. <question>
+  3. <question>
+  ## edge            # exactly ONE, must be framework-answerable
+  <question>
+  ## ground-truth
+  <the persona's real documented public stances — for the blind scorer only>
+  ## edge-note
+  <one line: why this edge is framework-answerable + that the persona never addressed it>
+"""
+import argparse
+import re
+import sys
+from pathlib import Path
+
+# --- grading thresholds (from nuwa fidelity-scorecard, + the F2' edge-honesty gate) ---
+GRADE_A = 85          # ship
+GRADE_B = 70          # acceptable with flagged weak spots
+EDGE_HONESTY_SHIP = 14   # /20 — a skill that fails edge-honesty is NOT ship-ready even if total is high
+REPRO_FLOOR = 0.60    # mental-model Jaccard overlap below this flags non-reproducible distillation
+
+DIMS = [
+    ("stance_consistency", 30),
+    ("style_recognizability", 20),
+    ("edge_honesty", 20),
+    ("source_transparency", 15),
+    ("structural_completeness", 15),
+]
+
+
+# --------------------------------------------------------------------------- spec validation
+def parse_spec(path: Path) -> dict:
+    text = path.read_text(encoding="utf-8")
+    sections = {}
+    cur = None
+    for line in text.splitlines():
+        m = re.match(r"^##\s+(\S.*)$", line)
+        if m:
+            cur = m.group(1).strip().lower().replace("-", "_").replace(" ", "_")
+            sections[cur] = []
+        elif cur:
+            sections[cur].append(line)
+    return {k: "\n".join(v).strip() for k, v in sections.items()}
+
+
+def validate_spec(path: Path) -> int:
+    s = parse_spec(path)
+    problems = []
+    if "known" not in s:
+        problems.append("missing '## known' section")
+    else:
+        known = [ln for ln in s["known"].splitlines() if re.match(r"^\s*\d+\.", ln.strip())]
+        if len(known) != 3:
+            problems.append(f"'## known' must have exactly 3 numbered questions, found {len(known)}")
+    if "edge" not in s or not s["edge"].strip():
+        problems.append("missing '## edge' section (exactly ONE novel edge question)")
+    if "ground_truth" not in s or len(s["ground_truth"]) < 50:
+        problems.append("missing or too-short '## ground-truth' (the blind scorer's reference)")
+    if "edge_note" not in s or "framework" not in s["edge_note"].lower():
+        problems.append("'## edge-note' must state why the edge is FRAMEWORK-answerable "
+                        "(fact-demanding edges give false passes — see f2-experiment/validation-conclusion.md)")
+    if problems:
+        print("❌ INVALID test spec:")
+        for p in problems:
+            print(f"  - {p}")
+        return 1
+    print("✅ test spec OK (3 known + 1 framework-answerable edge + ground-truth + edge-note)")
+    print(f"   edge question: {s['edge'].splitlines()[0][:90]}")
+    return 0
+
+
+# --------------------------------------------------------------------------- runbook emission
+RUNBOOK_HEADER = """# Fidelity runbook for {skill}
+
+Runtime-agnostic. Run the two steps below via THIS runtime's subagent mechanism, in order.
+Step 1 (answerer) MUST finish before Step 2 (scorer). The scorer is BLIND (ground-truth only,
+NO skill-file access) — F2 was refuted (skill-file access does not inflate scores) but the blind
+condition keeps the score independent of the author's design intent.
+
+## Step 1 — answerer (skill-only, NO web)
+Allowlist these files ONLY: <skill-dir>/SKILL.md + <skill-dir>/references/research/*
+Answer the 3 known + 1 edge question in character per the skill's Agentic Protocol.
+For the edge question, honor the skill's F2' third-category rule (flag framework-derived
+inference as inference, not established stance). Write answers to answers.md.
+
+## Step 2 — blind scorer (ground-truth ONLY, NO skill file)
+Allowlist: answers.md + rubric.md + ground-truth (from the test spec). Score 5 dims.
+Write scorecard to score-B.md.
+
+## Questions
+{questions}
+
+## Ground-truth (for blind scorer)
+{ground_truth}
+"""
+
+
+def emit_runbook(skill_dir: Path, spec_path: Path) -> int:
+    s = parse_spec(spec_path)
+    questions = "### known\n" + s.get("known", "") + "\n\n### edge (FRAMEWORK-answerable)\n" + s.get("edge", "")
+    out = RUNBOOK_HEADER.format(
+        skill=skill_dir.name,
+        questions=questions,
+        ground_truth=s.get("ground_truth", "(none)"),
+    )
+    dest = skill_dir / "fidelity-runbook.md"
+    dest.write_text(out, encoding="utf-8")
+    print(f"✅ wrote {dest}")
+    print("   next: dispatch Step 1 (answerer) then Step 2 (blind scorer) via your runtime.")
+    return 0
+
+
+# --------------------------------------------------------------------------- scorecard parsing
+def parse_score(path: Path) -> dict:
+    """Extract the 5 dimension scores from a scorer's markdown scorecard."""
+    text = path.read_text(encoding="utf-8")
+    found = {}
+    for name, _ in DIMS:
+        # match e.g. "| 1 Stance consistency | 26/30 |" or "| 3 Edge honesty | **6/20** |"
+        # [\s*]* tolerates bold (**..**) wrappers around the score.
+        pat = re.compile(rf"\|[^|]*{re.escape(name.split('_')[0])}[^|]*\|[\s*]*(\d+)\s*/\s*(\d+)", re.I)
+        m = pat.search(text)
+        if m:
+            found[name] = int(m.group(1))
+    total_m = re.search(r"\*\*TOTAL\*\*\s*\|\s*(\d+)\s*/\s*100", text, re.I)
+    if not total_m:
+        total_m = re.search(r"(\d+)\s*/\s*100", text)
+    total = int(total_m.group(1)) if total_m else sum(found.values())
+    return {"dims": found, "total": total}
+
+
+def report_score(path: Path) -> int:
+    r = parse_score(path)
+    if not r["dims"]:
+        print("❌ could not parse any dimension scores from", path)
+        return 1
+    print(f"Scores from {path}:")
+    for name, _max in DIMS:
+        v = r["dims"].get(name)
+        print(f"  {name:28} {v}/{_max}" if v is not None else f"  {name:28} (not found)")
+    print(f"  {'TOTAL':28} {r['total']}/100")
+    grade = grade_for(r["total"], r["dims"].get("edge_honesty", 0))
+    print(f"  grade: {grade}")
+    return 0
+
+
+def grade_for(total: int, edge: int) -> str:
+    if edge < EDGE_HONESTY_SHIP:
+        return f"NO-SHIP (edge-honesty {edge}/20 < {EDGE_HONESTY_SHIP} — fails the F2' gate regardless of total)"
+    if total >= GRADE_A:
+        return f"A (≥{GRADE_A}) — ship"
+    if total >= GRADE_B:
+        return f"B (≥{GRADE_B}) — acceptable with flagged weak spots"
+    return f"C (<{GRADE_B}) — re-distill"
+
+
+# --------------------------------------------------------------------------- reproducibility diff
+def extract_models(skill_md: Path) -> set:
+    """Extract mental-model names from a SKILL.md (### 模型N: / ### Model N: / ### N. Name)."""
+    text = skill_md.read_text(encoding="utf-8")
+    names = set()
+    # match model headings in either language; tolerate arabic OR Chinese numerals.
+    for m in re.finditer(r"^#{2,4}\s+(?:模型\s*[0-9一二三四五六七八九十]+\s*[:：]?\s*|Model\s*\d+\s*[:：]\s*|\d+\.\s*)(.+)$", text, re.M):
+        name = m.group(1).strip().split("(")[0].strip().split("（")[0].strip()
+        # drop parenthetical english + trailing junk
+        name = re.sub(r"\s+", " ", name)
+        if 2 <= len(name) <= 60:
+            names.add(name.lower())
+    return names
+
+
+def repro(a: Path, b: Path) -> int:
+    ma, mb = extract_models(a), extract_models(b)
+    if not ma or not mb:
+        print(f"❌ could not extract models (A={len(ma)}, B={len(mb)}). "
+              "Ensure SKILL.md uses '### 模型N: Name' / '### Model N: Name' headings.")
+        return 1
+    inter = ma & mb
+    union = ma | mb
+    jaccard = len(inter) / len(union) if union else 0.0
+    name_overlap = len(inter) / min(len(ma), len(mb)) if min(len(ma), len(mb)) else 0.0
+    print(f"Skill A models ({len(ma)}): {sorted(ma)}")
+    print(f"Skill B models ({len(mb)}): {sorted(mb)}")
+    print(f"intersection ({len(inter)}): {sorted(inter)}")
+    print(f"Jaccard overlap: {jaccard:.0%}   name-overlap/min: {name_overlap:.0%}")
+    if name_overlap < REPRO_FLOOR:
+        print(f"⚠️  NON-REPRODUCIBLE: overlap {name_overlap:.0%} < {REPRO_FLOOR:.0%} floor — "
+              "two distillations of the same persona diverge too much. Investigate Phase 2 subjectivity.")
+        return 2
+    print(f"✅ reproducible (overlap {name_overlap:.0%} ≥ {REPRO_FLOOR:.0%})")
+    return 0
+
+
+# --------------------------------------------------------------------------- CLI
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
+    sub = ap.add_subparsers(dest="cmd", required=True)
+    sub.add_parser("validate", help="validate a test-spec.md structure").add_argument("spec")
+    sub.add_parser("runbook", help="emit answerer+scorer prompts for a skill").add_argument("skill_dir")
+    [a.add_argument("spec") for a in [sub.choices["runbook"]]]
+    sub.add_parser("parse", help="extract 5-dim scores from a scorer's markdown").add_argument("score")
+    r = sub.add_parser("repro", help="diff mental models of two SKILL.md (reproducibility test)")
+    r.add_argument("skillA"); r.add_argument("skillB")
+    args = ap.parse_args()
+    if args.cmd == "validate":
+        return validate_spec(Path(args.spec))
+    if args.cmd == "runbook":
+        return emit_runbook(Path(args.skill_dir), Path(args.spec))
+    if args.cmd == "parse":
+        return report_score(Path(args.score))
+    if args.cmd == "repro":
+        return repro(Path(args.skillA), Path(args.skillB))
+    return 1
+
+
+if __name__ == "__main__":
+    sys.exit(main())