perf: cached repeated strict token lookups per input

[alyldas]

Summary

• Cache positive and negative normalized strict-token lookup results within one top-level invocation.
• Keep one- and two-token inputs on inline cache entries, then promote to a null-prototype dictionary bounded by the unique tokens in the current input.
• Reuse compiled pattern and taxonomy metadata while preserving per-occurrence boundary checks, UTF-16 ranges, and scanner early stop.
• Add regression coverage for repeated matches and misses, three-token promotion, boundary contexts, taxonomy filters, mutation between calls, and filter isolation.
• Extend the focused strict lookup benchmark with repeated-match timing and exact anchored pattern-evaluation counts.

Validation

• npm ci
• npm test -- tests/strict-token-index.spec.ts tests/api.spec.ts tests/scanner.spec.ts tests/runtime-literals.spec.ts tests/normalization.spec.ts tests/internal-rules.spec.ts
• npm run check
• npm run benchmark:strict-lookup with the same harness on origin/main and this branch
• npm run benchmark:profanity on origin/main and this branch with the same machine and command shape

Benchmark Evidence

Baseline origin/main:

• createProfanityFilter(): 7.1879 avg ms
• unique clean tokens: 1.0590 avg ms; 300 indexed evaluations per call
• repeated clean tokens: 0.8885 avg ms; 300 indexed evaluations per call
• repeated matches: 1.3303 avg ms; 300 indexed evaluations per call
• strict early match: 0.1475 avg ms
• strict late match after unique tokens: 0.3536 avg ms
• general long clean: 0.9494 avg ms
• general long late match with repeated clean text: 0.3700 avg ms

This branch:

• createProfanityFilter(): 7.2239 avg ms
• unique clean tokens: 1.1150 avg ms; 300 indexed evaluations per call
• repeated clean tokens: 0.6630 avg ms; 1 indexed evaluation per call
• repeated matches: 1.2902 avg ms; 1 indexed evaluation per call
• strict early match: 0.1519 avg ms
• strict late match after unique tokens: 0.4022 avg ms
• general long clean: 0.8792 avg ms
• general long late match with repeated clean text: 0.1680 avg ms

Repeated clean lookup improved by about 25.4%, repeated match analysis by about 3.0%, general long clean by about 7.4%, and the repeated long-late fixture by about 54.6%. The unique-token negative control measured about 5.3% overhead, with no change to its 300 required indexed evaluations; the cache is lazy for up to two distinct tokens and remains bounded by the current input.

Compatibility Notes

• Public APIs, exports, scanner inputs, tokenization, normalization strategies, and dictionary formats are unchanged.
• Strict rule precedence, taxonomy metadata, UTF-16 analyzer ranges, scanner code-point ranges, and occurrence-specific boundary decisions remain compatible.
• Positive and negative entries live only inside one strict scan, so runtime mutation between calls and multiple filter instances cannot share stale cache state.
• Boolean checks retain early stop, while analyze, censor, and streaming scans reuse lookup results without skipping occurrence-specific ranges.
• Release Please PR chore(main): release 0.16.2 #175 remains open and is reserved for the single milestone release.

Closes #172

No publish, no merge, no tag/release.