deepnsm-v2: keep counted lexical evidence beside routing

.claude/board/entries/2026-09-26-deepnsm-v2-lexical-evidence-survives-routing.md

@@ -0,0 +1,16 @@
+# 2026-09-26 — DeepNSM-v2 keeps counted lexical evidence beside routing
+
+**Status:** MEASURED · DONE — `crates/deepnsm-v2/src/lexical.rs`
+
+## The loss point
+`PaletteVocab::from_frequency_ranked` (`vocab.rs`) admits surface strings, first wins. The counts and PoS behind the ranking never reach v2 storage. `examples/bible_wave.rs::load_pos` repeats the loss for PoS: `word -> one Pos`, counts dropped.
+
+## What landed
+`LexicalEvidence`, built against an existing `PaletteVocab` and never changing it. One `WordId` owns N `LexicalReading { pos: PosCode, lemma: Option<LemmaRef>, form_count: Option<u64> }`. `LemmaEntry { source_key, lemma, pos, count }` is keyed by the source's `lemRank`. Queries: `surface_count`, `surface_pos_count`, `lemma_pos_count`, `lemma_count`. All counts are `Option<u64>`: an empty field is unknown, a literal `0` is zero. Conflicting lemma entries and duplicate readings are refused, not first-wins. Loader: `load_word_forms_csv` (exact header `lemRank,lemma,PoS,lemFreq,wordFreq,word`).
+
+## Measured on the committed COCA `word_forms.csv`
+11,460 rows → 11,456 readings + 4 empty-surface rows (reported), 0 unrouted when the vocab covers every surface, 5,050 lemma entries (= distinct `lemRank`). 1,165 surfaces carry more than one reading, e.g. `record` n 120,048 / v 13,014 under lemma totals 187,057 / 51,375.
+
+## Open
+- The trained Cam96 path (`bible_vocab.txt`) is surface-only: no counts, no PoS. KJV routing gets evidence only when a caller joins a counted source to it.
+- `academic_20k.csv` (surface × PoS × COCA-All/Acad, no lemma) has no loader. It holds 3 same-(word, PoS) row pairs with different counts, so a loader must decide whether those are disjoint before summing. The builder refuses a second (word, PoS, no-lemma) reading as `DuplicateReading` until that is decided.

.claude/board/entries/README.md

@@ -25,10 +25,11 @@ index row, (3) no duplicate entry id. Checks 1 and 2 are deliberately
 opposite directions; the stranding this convention prevents shows up in
 exactly one of them, never both.
 
-174 entries, 2026-08-06 .. 2026-09-25.
+175 entries, 2026-08-06 .. 2026-09-26.
 
 | date | entry id | finding | file |
 |---|---|---|---|
+| 2026-09-26 | `deepnsm-v2-lexical-evidence-survives-routing` |  | [2026-09-26-deepnsm-v2-lexical-evidence-survives-routing.md](2026-09-26-deepnsm-v2-lexical-evidence-survives-routing.md) |
 | 2026-09-25 | `window-scheduling-and-two-level-ternlog` |  | [2026-09-25-window-scheduling-and-two-level-ternlog.md](2026-09-25-window-scheduling-and-two-level-ternlog.md) |
 | 2026-09-25 | `tile-knee-and-compiled-lowering` |  | [2026-09-25-tile-knee-and-compiled-lowering.md](2026-09-25-tile-knee-and-compiled-lowering.md) |
 | 2026-09-25 | `tern3-six-view-window` |  | [2026-09-25-tern3-six-view-window.md](2026-09-25-tern3-six-view-window.md) |

crates/deepnsm-v2/data/README.md

@@ -15,3 +15,12 @@ curl -L -o bible_vocab.txt    "$BASE/bible_vocab.txt"      # frequency-ranked vo
 Provenance + held-out metrics: `../probes/README.md` §4 (producer scripts:
 `../probes/{embed_bible_vocab,train_codebook}.py`). Loaded by
 `deepnsm_v2::codebook::{load_cam96_space, load_cam96_codes}`.
+
+## Lexical evidence (counts + PoS) — not in the Cam96 release
+
+`bible_vocab.txt` carries surface forms only, in rank order: no counts, no
+part of speech. Counted evidence is read separately by
+`deepnsm_v2::lexical::load_word_forms_csv`, whose input is COCA
+`word_forms.csv` (`lemRank,lemma,PoS,lemFreq,wordFreq,word`, committed at
+`crates/deepnsm/word_frequency/`). It is stored beside the routing
+`PaletteVocab`, never inside it and never inside the Cam96 codes.