Add GrammarAwareRanker for best-of-N self-consistency

[piroz]

Implements the next-step "B" from the research spike: grammar-aware best-of-N reranking to close the small-LLM 1-shot miss gap, while staying logit-free.

Problem

When selfConsistency.n > 1 and none of the N candidates fully validates, the old "first-valid wins" picker silently degraded to "first candidate wins" — the rest of the self-consistency signal was thrown away. The error feedback fed into the retry loop was therefore whichever sample happened to be sampled first, not whichever sample got closest to a valid program.

Approach

• ValidationResult gains an optional errorOffset so a validator can report how far it parsed before giving up.
• Ranker.rank gains a context?: { validations } parameter so a ranker can score off that progress without re-parsing.
• GrammarAwareRanker (new) scores valid → 0, else max(1, len - errorOffset) — lower wins, so:
  • any fully-valid candidate beats every invalid one;
  • among invalid candidates, the deepest parse beats a shallow one.
• The picker (src/noroshi.ts::pick) is now async, threads validations through, and the out-of-retries return surfaces the picked output (was dropping the ranker's choice into lastCandidates[0]).

Why this is worth doing now

• Directly attacks noroshi's biggest weakness: small models (1B-3B) frequently miss on the first try.
• Stays inside the prompt-side contract (no logit access required) so every adapter — FetchAdapter, WebLLMAdapter, future ChromePromptAdapter — gets the benefit equally.
• The deepest-failed-parse heuristic feeds the retry loop the most specific error message, which compounds with retry.includeErrorInPrompt — basically validity-aware sampling in the spirit of CRANE / GAD without needing constrained decoding.

Tests

New suite test-ranker.ts (6 cases):

• Direct rank() ordering: valid (0) < deep-fail (1) < shallow-fail (24).
• End-to-end through generate():
  • all 3 candidates invalid → deepest parse wins
  • mixed set → valid candidate wins regardless of position
  • no-ranker baseline → first-valid still wins (regression)

Combined with the existing suites (test.ts 9 / test-fetch.ts 8 / test-safety.ts 15), npm test now runs 38 assertions, all green locally and on Node 20 + 22 (waiting on CI).

Files

• src/types.ts — ValidationResult.errorOffset?, Ranker.rank(..., context?)
• src/rankers/grammar.ts (new) — GrammarAwareRanker
• src/index.ts — exports GrammarAwareRanker
• src/noroshi.ts — async pick, lastOutput fallback fix
• examples/creative-coding-p5js/validator.ts — lex + parser emit errorOffset
• examples/creative-coding-p5js/harness.mjs — mirrors the above for the browser path
• examples/creative-coding-p5js/test-ranker.ts (new) — 6 cases
• package.json — npm test runs the new suite alongside the others