v5.0.0 — Audit-driven fixes + progressive-disclosure restructure

Ships v4.3.0 (audit-driven fixes) and v5.0.0 (restructure) together. Full details in CHANGELOG.md.

v5.0.0 · Progressive-disclosure restructure 渐进式披露重构

• SKILL.md body ~2,000 → ~840 lines — per-activation context cost drops from ~35k to ~18k tokens. 主文件瘦身一半以上，每次激活的上下文成本减半
• 9 new bilingual reference pairs, loaded on demand: disciplines · modes-prewriting (H/I/J) · mode-c-drafting · mode-d-adversarial · mode-e-bottleneck · mode-f-revision · modes-submission (G/K) · deep-style · multilingual-writing. Sections moved verbatim; every mode keeps a decision-carrying stub in the core file. 九对按需加载的双语手册；核心文件中每个模式保留带硬约束的存根
• Selective Loading Guide rewritten as the actual router (task → which file to Read). 按需加载指南重写为真正的路由表

v4.3.0 · Audit-driven fixes 审计驱动修复

A 31-agent QA audit (14 mode simulations, 4 static audits, 7-channel market scan, 5-judge panel) → every confirmed defect fixed:

• Routing: frontmatter description 1,727 → 992 chars (was 69% over the 1,024-char Agent Skills spec limit — the most aggressive triggers never reached the router on truncating platforms) 触发描述压缩至规范限内
• Toolchain integrity: citation-verify.py no longer rubber-stamps hallucinated citations (FOUND now prints the matched title for human confirmation; new ERROR verdict; Crossref→OpenAlex cascade); ai-trace-scan.sh directory mode actually works now (previously printed a false "✅ no AI traces") 反幻觉脚本的两处目的反转缺陷修复
• Mode upgrades: Mode E first-response protocol + bottleneck typology + capability boundary; Mode L four-way triage + response-letter/修改说明 workflow + traceability matrix; Mode K policy-never-from-memory + real Template C + anti-whitewashing rules; Mode D evidence contract + review-the-review + two-stage option
• New: references/style-profile-template bilingual pair (the "constitution of voice" finally has a template); rule-precedence + global flagged-diff rules; minimal-start protocol; 58-assertion test suite + CI (bilingual heading-parity across 18 file pairs); release checklist
• 12 mechanisms adapted from the strongest peer tools (evidence contract, review-the-review guardrail, rhetorical-action menu, process-ledger disclosure, over-imitation reverse check, …)

Compatibility 兼容性: existing _writing-config/ / _meta/ project files continue to work unchanged. The skill has required multi-file support since v4.x; packaging shape is unchanged. 既有项目文件无需迁移。