feat(vocab): learned spell-correction + user dictionary — context, not dictionary matching (task #13)

[mdheller]

The dictionary should be learned, not matched

A static wordlist flags every domain term (epistemiclevel, srcos, governedloop) as a misspelling and corrects it away — the estate's own vocabulary treated as errors. So what counts as a correct word must be learned from context. For each UNKNOWN token tools/learned_dictionary.py decides from a skip-gram word-sense predictor:

• learn (add to user dictionary) — recurs with a COHERENT context (context windows cluster into one stable sense): a real term the dictionary hadn't seen.
• correct (to a known word) — rare AND spelling-near AND sense-near (skip-gram cosine). Sense, not edit distance alone, picks the target — a token spelled near a known word but used in a different sense is not auto-corrected.
• unknown — neither coherent nor sense-close: left for a human.

Predictor = count-based skip-gram: PPMI over a co-occurrence window + truncated SVD (Levy-Goldberg: SGNS factorises shifted PPMI — same word-sense family). No wordlist decides correctness; context does. Fail-closed: every decision is a proposal (add / correct-to), never a silent rewrite.

Teeth (make validate-learned-dictionary)

• epistemiclevel (novel, recurs coherently) → learned, not corrected away
• reciept → corrected to receipt by learned sense (carries senseSim)
• qwzptl (garbage) → unknown, fail-closed
• a learned term is never auto-corrected away

Same doctrine as the stopword analysis (#260/#261) and glossary currency (#255): replace static membership tests with learned, context-driven predictors. Uses numpy (installed in the make recipe).