Add Islamic & manuscript content extensions

[SherlockianAsh]

PR: Add Islamic & Manuscript Content Extensions

Summary

Adds Malaysian Islamic / Sufi / kitab-translation register support to Manusiawi. The current skill defaults to "natural Malaysian writing" (casual / professional register), which produces false-positive AI flags when applied to religious content where formal Arabic-derived vocabulary and conventional kitab structures ARE the natural register.

This PR extends Manusiawi with three new reference files plus SKILL.md adjustments. The defaults remain unchanged — the new behaviours activate when religious-content signals are detected (≥3 Islamic terms per 500 words) or via explicit register flag.

Motivation

Built and tested against real Malaysian Islamic content production:

• Classical Arabic kitab → BM translation (Or_7066 by Sheikh Sulayman Zuhdi, Kifayah al-Ghulam by Sheikh Ismail al-Minangkabawi, Al-Bahjah As-Sunniyyah, Al-Hadaiq ul-Wardiyya).
• Naqshbandiyya-Khalidiyya tariqat documentation (in-school suluk programs, ratib publications).
• Modern dakwah and pondok education materials.

Without the extension, running Manusiawi on these documents flags legitimate religious vocabulary (tasawuf, hakikat, tazkiyah) as Pattern 6 (AI vocabulary), legitimate hadith narration formulas (telah berkata Imam Bukhari) as Pattern 11 (excessive telah di- passive), and conventional kitab closings (semoga Allah merahmati) as Pattern 24 (generic positive conclusions). The false-positive rate makes the skill unusable for this register.

The extension preserves Manusiawi's anti-AI rigor while making it correctly recognize religious content conventions.

What's Added

references/islamic-terminology.md (new)

Whitelist of Islamic / Sufi technical vocabulary that must NOT be flagged as Pattern 6 (AI vocabulary) when religious register is detected. Categorized by domain (aqidah, fiqh, tasawuf, akhlak, devotional forms, Sufi orders, honorifics, manuscript genre). Includes consistency rules for prophetic honorifics (SAW / ﷺ / s.a.w.) within a document.

Edge case rule: even in religious register, holistik / komprehensif / ekosistem STILL flag — those are AI fluff regardless of domain. The extension is precise: it whitelists technical religious terms, not all heavy vocabulary.

references/islamic-transliteration.md (new)

Standard Malaysian Islamic transliteration table per DBP (Dewan Bahasa dan Pustaka) Pedoman Transliterasi Huruf Arab ke Huruf Rumi. Covers daily prayer terms, Hijri month names, prophetic and other honorifics, Allah-related forms, and kitab title conventions.

Addresses recurring drift cases (sholat → solat, jum'at → jumaat, dzikir → zikir, ramadhan → ramadan) and consistency rules within documents.

references/indonesian-words-islamic-additions.md (new)

Extends indonesian-words.md with religious-context Indonesian intrusions:

• Religious institutions: pesantren → pondok, kyai → tok guru, santri → pelajar pondok
• Ritual terms: sholat → solat, dzikir → zikir, kurban → korban
• Honorifics: pak ustadz → ustaz, dakwah jargon (antum, akhi/ukhti) → BM equivalents
• Mosque/community: mushalla → surau, tausiah → tazkirah/nasihat
• Sufi lexicon drift: thoriqoh → tariqat, sholawat → selawat
• Common Indonesian dakwah idioms with Malaysian replacements

This file is structured for either standalone reference loading OR merging into indonesian-words.md directly. Maintainer's preference.

SKILL.md (additions)

Two new sections proposed: "Mode: Religious & Manuscript Content" and "Submode: Kitab Translation". They document:

• Detection signals for auto-activating religious mode (3+ Islamic terms per 500 words OR explicit user request)
• Adjusted pattern application for religious register (which patterns to flag, which to skip, which to apply with modifications)
• Conventional kitab register patterns to preserve (passive narration, stock phrases, Adapun/Maka/Tiada conventions)
• Explicit "do NOT humanize" list (Quranic verses, hadith proper, du'a, wirid/hizib, direct Arabic transliterations) — Mei surrounding commentary may be humanized but never the religious source texts.

Defaults unchanged: skill behaves identically for non-religious content. Religious adjustments apply only when register is detected or flagged.

Test Cases (informal)

Tested manually against samples from the four kitab translation projects. Without extension: false-positive rate ~40% (legitimate Sufi terms flagged). With extension: false-positive rate ~3% (only genuinely AI-flavoured passages flagged).

Indonesian intrusion detection caught additional terms not in the original indonesian-words.md (pesantren, dzikir, tausiah, pak ustadz) when applied to dakwah-style Indonesian-leaked AI output.

Risk / Backward Compatibility

• No existing pattern is removed — only register-conditional adjustments to flag intensity.
• Default register behaviour unchanged — non-religious content sees zero behaviour change.
• New whitelist is additive — adds terms that should never be flagged, doesn't remove or weaken anti-AI flagging.
• New transliteration file is reference-only — informational, doesn't change skill logic unless explicitly invoked.

Possible Reviewer Concerns

1. Scope creep — This extension is Malaysian Islamic specific. Would Indonesian Islamic users want a parallel extension? Possible, but out of scope here. The structure of the new files is parallel-friendly: future Indonesian Islamic extension could mirror this one.
2. Whitelist drift — As Islamic vocabulary evolves and AI vocabulary mutates, the boundary between "technical religious term" and "AI fluff" may shift. The PR proposes the boundary as of 2026; future updates should track DBP and AI-vocabulary drift.
3. Detection heuristic strictness — "3 Islamic terms per 500 words" is a starting heuristic. Field-tuning may be needed.

Out of Scope

• Indonesian Islamic register extension (would mirror this PR for Indonesian audiences).
• Christian / Hindu / Buddhist Malaysian religious content (each has own register conventions).
• Arabic-only manuscript humanization (skill is BM/EN, not Arabic).

---

If this PR is too broad, the maintainer can cherry-pick: islamic-terminology.md alone is the minimum-viable improvement and resolves the worst false-positive rate.