revert(lexicon): remove multi-word phrase support

[elecnix]

Reverts the phrase half of #45. Reopens #40.

Why

Phrases do not work, and three attempts did not change that.

Over the real corpus they were 84% of all adjudications (65,792 of 78,437 entries) and 75% of all hits (28,179 against the word lexicon's 9,350), and the output was noise.

Tightening the prompt, then adding a deterministic rule suppressing any phrase whose component word had already fired, cut the volume hard — flag rate 6.66% → 1.65%, hits 28,179 → 1,525. But what survived was still not idiomatic:

praise      | accepted or          frustration | after another
praise      | accepted feedback    frustration | again for
frustration | aach wbaagb          frustration | admin's mock

top hits: first quality x111 · end your x69 · re check x43 · an explicit x26

The premise was wrong, not the tuning. Adjacent word pairs in running prose are overwhelmingly not fixed expressions, and no judge reliably finds the rare real ones among tens of thousands of candidates. Lowering the noise floor does not turn a bad candidate set into a good one.

What remains

The parts that measurably work:

• single-word judgement — 3.6% flag rate, producing stuck, pedantically, salty, liar, hell, masterful, 🍾
• paralinguistic markers — shouting, ???, elongation; no model required

What is lost

laisse tomber and trop lent — the multilingual idioms that motivated #40. They are real, and #40 is reopened. A future attempt needs a different mechanism: a curated idiom seed list, or statistical n-gram selection, rather than per-pair LLM adjudication of every adjacent word pair. DESIGN.md records the withdrawal and the measurements so the next attempt does not restart from the same assumption.

Test plan

• npm test — 424 passing (phrase-specific tests removed).
• node --import tsx test/integration/test-commands.ts — 21 passing.
• npx tsc --noEmit clean.

🤖 Generated with Claude Code