feat(engine): understand text as people type it — rasm-uniqueness proof (D-015) + unmarked-quote detection (E-057/E-058)

[MoTechSys]

Why

Supervisor feedback (أ. أماسي, 2026-10-05): typed «إنما الأعمال بالنيات» alone → zero matches; added «قال رسول الله …» → found. «The judges will say the site does not work.» أ. أفنان: trust is the value — test diverse real inputs, measure clarity and stability, not fixed examples.

Two root causes, both fixed here, both measured:

1. Unmarked short quotes were invisible — the detector required 6 content words and treated «الله / لا / من / إن» as stop-words.
2. Hamza-less typing was treated as an error — «قل هو الله احد» → «اختلاف إملائي — يحتاج مراجعة».

Before → After (full index, offline mock provider)

Input (no marker, as people type)	Before	After
إنما الأعمال بالنيات (supervisor's input)	❌ no quote	found · 3 positions
انما الاعمال بالنيات	❌ no quote	found + rasm_completed=[إنما, الأعمال]
قل هو الله احد	❌ no quote	found · 112:1 + completed أحد
ان الله مع الصابرين	❌ no quote	found · 2:153 / 8:46 + completed إن
لا ضرر ولا ضرار · إنما بعثت لأتمم مكارم الأخلاق · خيركم من تعلم القرآن وعلمه · من غشنا فليس منا · الكلمة الطيبة صدقة	❌	found
واذا سالك عبادي عني فاني قريب	❌	found · 2:186 + 3 completions
ان الله علي كل شيء قدير (judges' trap)	needs_review / orthographic_difference	needs_review / rasm_ambiguous — shows إنّ… (8) / أنّ… (3), never found
ذهبت اليوم إلى السوق… · حب الوطن من الإيمان · الدين المعاملة	no quote	no quote (unchanged)
حدثنا عبد الله بن يوسف قال أخبرنا مالك عن نافع عن ابن عمر	found (record isnad)	unchanged — rare tail «نافع عن ابن عمر» never anchors

What

• D-015 rasm-uniqueness proof (I18/I19) — match/rasm.py: classify each differing token completion (unwritten hamza/ى/ة) vs swap (user wrote a different mark) vs other; census the strict spelling across every corpus position. One spelling + all completions → found + rasm_completed (letters listed with char spans). Else → needs_review/rasm_ambiguous + alternatives with counts. V7 re-derives the proof from the Store alone. QuoteResult.rasm exposes it.
• E-057 whole-ayah anchor — a run that is a complete ayah is a quote regardless of stop-words.
• E-058 phrase-rarity anchor — a run of ≥ 3 tokens occurring ≤ 60× verbatim in the corpus is a quote (quotations: 2–27×; formulaic prose: 105–92 083× — the band between is empty); runs extend backwards over leading particles; isnad guard (chain-shaped runs never anchor).
• Composer hint (ar/en): «اكتب النص كما هو — لا يلزم قال تعالى… ولا الهمزات ولا التشكيل».
• Eval: category-B expectations resolved by construction from the corpus (materialize.rasm_expectation), same records.
• Docs: SAFETY (I18/I19, V7), DECISIONS (D-015, E-057, E-058), CHANGELOG, messages ar/en.

Measured

• pytest 344 (was 304): +25 test_rasm.py, +15 test_as_people_write.py · ruff + mypy strict clean · lexicon 658 strings / 0 forbidden · vitest 30/30 · build ok
• make smoke: 8 canonical + 7 unmarked cases (incl. the supervisor's exact input) → SMOKE OK
• make eval-full: 150/150 · unsafe 0 · variance 0 · false alarms 0/500 (full index); make eval (fixture): 150/150
• IslamicEval 1B: 78.54 %, false confirmations 2 (unchanged)
• Tanzil census: 87.7 % of token occurrences have one Mushaf spelling for their bare form; the 12.3 % = إن/أن · على/علي · إلا/ألا → stay in review

Known limit (documented)

• 2-token sayings («الدين النصيحة») still need a marker: below RARE_MIN_TOKENS=3 the false-alarm risk is unmeasured.

Follow-ups (next PRs)

• Frontend: render rasm.completions (highlight completed letters) and rasm.alternatives (choose-one); card catalog / flip view; HTML/MD/PDF export.
• bootstrap.sh builds the frontend (fresh clone → / is 404 today).
• /settings shows "configured ✅" when protected.