relationship: privacy screening and consent gates; work: non-IT role coverage

[SCM-Consultant-Ayaan]

Summary / 摘要
Two related improvements to the distillation pipeline: consent gating plus a privacy screener for the relationship preset, and role coverage for the work track so it stops assuming every subject is an IT worker. Both operationalize governance and quality points the technical report leaves as open work.
为蒸馏流程补了两块：relationship 家族的同意关卡与隐私筛查工具，以及 work 轨道的职位覆盖（不再默认对象是 IT 从业者）。两者都把技术报告中留作 open work 的治理与质量要求落到可执行的代码。
Changes / 变更

1. relationship — privacy screening + explicit consent gates
   New tools/privacy_screen.py (655 lines). Scans collected raw material under knowledge/{slug}/ before it reaches the analyzers and writes a structured privacy_report.json with an overall CLEAR / NEEDS_REVIEW status.
   Detects phone numbers, emails, Korean RRNs and Chinese national IDs (both checksum-validated), bank accounts (keyword-gated to avoid firing on order numbers), addresses, and Luhn-valid card numbers.
   Flags third-party speakers — participants other than the target and the user — since personal chat exports routinely carry other people's messages.
   Flags sensitive categories: health, sexual content, minors, financial, political/religious.
   Never drops or rewrites source material. Identifier values are masked in the report; third-party display names are shown in the clear so the user can tell who they are being asked about.
   Emits an explicit "not covered by this screen" list on every run, so a CLEAR status is not read as a guarantee.
   SKILL.md — new screening step between source import (Step 2) and analysis (Step 3) for the relationship flow. Analysis cannot start while the status is NEEDS_REVIEW and unresolved; the user decides keep / redact / exclude per flagged category.
   SKILL.md — separate, explicit confirmation gates before writing files, before host installation, and before any gallery publication. Gallery publication is its own gate rather than being implied by the write-files answer. Silence is not treated as consent.
   prompts/relationship/* — correction handling now shows the concrete patch and waits for approval instead of auto-applying to persona.md.
   KakaoTalk export support in the speaker detector: the desktop format (2024년 1월 2일 오후 3:15, 이름 : …) and the bracketed format ([이름] [오후 3:20] …) previously matched no speaker pattern at all, so the most common Korean chat export silently reported zero third parties.
2. work track — non-IT role coverage
   prompts/work_analyzer.md — role-specific extraction previously covered six IT roles only (backend, frontend, ML, PM, designer, data analyst). Anyone outside that list fell through to the four generic dimensions. Added:
   A rule for roles absent from the list: derive the subject's 3–5 recurring work types, then extract decision criteria, action order, exceptions, escalation thresholds, and red lines. Those four hold for any profession and beat borrowing mismatched IT dimensions.
   A supply-chain / SCM profile covering forecasting (baseline method, manual-override conditions, accuracy tolerances), inventory trade-offs (stockout vs obsolescence preference, safety-stock logic), sourcing (supplier tiering, concession order across price / payment terms / MOQ / lead time, delay escalation thresholds), logistics, exception handling, and S&OP.
   A consulting / professional-services profile covering diagnosis order, deliverable standards, and client communication boundaries.
   prompts/work_builder.md — the render template hardcoded 技术栈 / 代码风格 / 接口设计 / Code Review headings, which produce empty or nonsensical sections for non-engineering subjects. Headings are now role-adaptive, the builder is told to delete inapplicable ones rather than leave shells, and the concreteness examples cover SCM and consulting alongside engineering.
3. Output language fix
   prompts/work_analyzer.md and prompts/persona_analyzer.md both hardcoded 语言：中文, so a Korean or English user received Chinese analysis regardless of the language they used. Both now follow the user's language, which is what SKILL.md already requires for meta.classification.language.
   Scope note: no changes to the colleague or celebrity flows beyond the shared work_* prompts and the two analyzer language lines described above.
   Motivation / 动机
   Section 9 of the technical report states that responsible deployment "requires explicit participation, scoped source collection, access controls, retention limits, and non-mandatory use," and that gallery publication should stay opt-in with review. The relationship preset is the family where those requirements bite hardest — its evidence is private interpersonal history, and Section 4.4 already frames deletion, correction, and non-public defaults as first-order artifact requirements. Before this change those properties existed as prose in the prompt rather than as steps that block progress, and nothing inspected the collected material for third parties or identifiers at all. This PR turns them into gates the flow cannot skip.
   The work track change addresses a narrower gap. The report presents the colleague setting as the primary and most controllable case, with review checklists and escalation thresholds as the useful payload. That payload is exactly what the analyzer failed to elicit outside software roles: for an SCM planner or a consultant, the questions that matter (what triggers an escalation, which concession comes first, what gets overridden) were never asked, and the render template then labelled their output with engineering headings.
   No linked issue — happy to open one first if maintainers prefer that order.
   Testing / 测试
   [x] python -m unittest discover -s tests -p 'test_*.py' passed — 61 tests, OK (26 pre-existing, 35 new in tests/test_privacy_screen.py)
   [x] python -m compileall tools/ passed
   [x] Manually tested: ran privacy_screen.py against synthetic KakaoTalk exports in both desktop and bracketed formats, confirming third-party speakers are detected with message counts, phone numbers are masked in output, and the limitations list is printed. Also confirmed a clean transcript reports CLEAR.
   New tests cover each detector category, the masking guarantee (asserting raw RRN / card / email values never appear in the report), CLEAR vs NEEDS_REVIEW status, both KakaoTalk formats, the comma-separated timestamp case, and false-positive guards (bare 6-digit codes are not RRNs, order numbers are not bank accounts, Note: prose is not a speaker, a single-line label is not a speaker).
   No new runtime dependencies — privacy_screen.py uses only the standard library.
   Checklist / 检查清单
   [x] I read CONTRIBUTING.md
   [x] Docs updated if behavior or usage changed (README / SKILL.md / INSTALL.md) — SKILL.md updated in both the Chinese and # English Version halves; README.md and docs/lang/README_ZH.md note the new screening gate and its limitations
   [x] No secrets, tokens, or personal data committed — test fixtures are synthetic values generated to satisfy each checksum
   [x] New dependencies added to requirements.txt (if any) — none needed
   [x] Tests added/updated for new functionality
   Notes for reviewers
   Happy to split this. The two concerns are independent: privacy gating for relationship (commits 1–2) and role coverage for the work track (commit 3). If you would rather review or merge them separately, say so and I will open two PRs.
   The one deliberate asymmetry worth scrutinizing: identifier values are masked in privacy_report.json, but third-party display names are not. Masking a two-character Korean name yields **, which makes the "should this person's messages be excluded?" question unanswerable. The name already sits in the source file beside the report, and the report never leaves the machine, so this trades no new exposure for a usable gate — but it is a judgment call, and the reasoning is recorded in the code and in the report's limitations list.
   Detection is pattern-based and deliberately conservative. It will miss paraphrased sensitive topics, third parties mentioned only inside message bodies, and anything in images or audio. The screener says so on every run rather than implying completeness.