Improve English tokenization: stemming, larger stop word set, fix HEX false positives

[chigichan24]

Why

• The English path in scripts/knowledge-graph/tokenizer.ts handles whitespace, kebab / snake / CamelCase splits, and lowercasing — but is otherwise quite naive.
• HEX_PATTERN = /^[0-9a-f]{6,}$/i (scripts/knowledge-graph/constants.ts:39) matches plain English words like decade, facade, effect, defaced. They get dropped as "noise" even though they are normal vocabulary, throwing away signal.
• The hand-rolled stop word list (constants.ts:7-34, ~100 entries) is significantly smaller than industry-standard sets (NLTK ~180, spaCy ~325). Real session prose contains many high-frequency words that slip through and dilute IDF weights.
• No stemming: run / running / runs, test / tested / testing, fix / fixed / fixing end up as separate vocabulary entries, fragmenting IDF and weakening cosine similarity between clearly related sessions.

What

• Fix the HEX_PATTERN false positive: a real hex literal must contain at least one digit (otherwise it is a normal English word built from a-f).
• Add Porter (or Snowball) stemming after CamelCase split, before stop-word filtering.
• Expand STOP_WORDS to roughly NLTK parity (~180 entries).
• Add tests that:
  • assert stemming collapses run/running/runs
  • assert decade, facade, effect are not filtered as noise
  • assert the new stop words are dropped

How

• Porter stemmer: ~70 LoC of pure logic, ship inline (no deps) or use a single-purpose library that has no transitive deps.
• HEX_PATTERN: change to /^(?=[0-9a-f]*[0-9])[0-9a-f]{6,}$/i (require ≥1 digit). Sanity-check against real session log fixtures.
• Be careful at the Japanese boundary: stemming must not apply to tokens containing CJK characters.

Notes

• Independent from the Japanese tokenizer issue. Can ship in either order. The TF-IDF modernization issue benefits from this landing first.