docs: v11c card fixes — tatoeba ladder, 12 diverse lanes (algo=6), literary lane, 32k cap, dataset link (was: stale data_mix4 pointer)

README.md

@@ -106,7 +106,7 @@ printf 'Қазақша\nHello\n' | spellman detect --lines               # kaz /
 |---|---|
 | held-out mix (352,470 rows, pristine test) | 98.56% |
 | Tatoeba (37,051, out-of-domain) | 98.78% |
-| Tatoeba granularity ladder (word / pair / triple) | 66.7 / 86.2 / 93.3 |
+| Tatoeba granularity ladder (word / pair / triple) | 70.6 / 88.2 / 94.5 |
 | wild Russian tweets (2,606, label-audited; 61% @mentions, 20% URLs) | 94.47% |
 | COSMUS gold-labeled wild Russian (2,808, unseen) | 97.15% |
 | short-utterance referee (orthography-certified, ≤19 chars) | 91.29% |
@@ -124,13 +124,19 @@ set still leads clean out-of-domain sentences (99.25 vs 98.78).
 
 ## Training data
 
-FineWeb-2 per-language configs, ~104k Tatoeba sentences (CC BY 2.0 FR),
-per-language community/parallel corpora, a wild social-media lane (1.85M
-Ukrainian tweets, Telegram Uzbek, Chuvash CC0 web, Kazakh reviews,
-Leipzig community crawls, cardiffnlp multilingual tweets, sentiment140,
-2025 German-politics tweets), a verified short-utterance lane (3-19 char
-wild rows under 2-of-3 judge consensus and orthographic twin rules), and
-lexically-diverse rus/ukr sets selected by greedy lemma coverage
-(pymorphy3; 93% of the top-15k lemmas in 14k rows). Every source
-LID-hygiene-filtered with twin-language protection; the exact recipe is
-recorded in the repository (data_mix4/manifest.json).
+The full training mix (1,535,170 rows, 26 languages) is published at
+[vpermilp/spellman](https://huggingface.co/datasets/vpermilp/spellman)
+(dataset repo type) — the byte-exact recipe lives in its `manifest.json`.
+
+Built from FineWeb-2 per-language configs (32k rows/language cap),
+~104k Tatoeba sentences (CC BY 2.0 FR), per-language community/parallel
+corpora, a wild social-media lane (1.85M Ukrainian tweets, Telegram
+Uzbek, Chuvash CC0 web, Kazakh reviews, Leipzig community crawls,
+cardiffnlp multilingual tweets, sentiment140, 2025 German-politics
+tweets), a verified short-utterance lane (3-19 char wild rows under
+2-of-3 judge consensus and orthographic twin rules), a literary Russian
+lane (wikisource classic prose, 6k rows), and 12 lexically-diverse lanes
+selected by banded lemma coverage (algo=6) over wild pools, with real
+normalizers — pymorphy3 for rus/ukr, Stanza for bel/bul/srp/kaz/kir,
+Apertium for mkd, corpus-derived prefix stems for the agglutinative
+set. Every source is LID-hygiene-filtered with twin-language protection.