Remove per-token allocation and sorting from strict lookup

[alyldas]

Context

The strict token lookup currently allocates a fallback array, a Set of seen rule orders, a lookup-character Set, and then sorts candidates for every token. This work is avoidable because compiled patterns and buckets already have stable rule order.

Proposed Implementation

• Preserve rule order in compilation-time token-index structures.
• Replace per-token candidate array cloning, de-duplication Set creation, and sorting with an allocation-bounded ordered lookup.
• Evaluate fallback and indexed buckets without changing first-match precedence.
• Prefer numeric pattern ids and reusable per-call scratch state, premerged buckets, or another approach with explicit reentrancy and mutation semantics.
• Keep symbol-token and phrase behavior unchanged unless profiling identifies the same allocation pattern there.
• Add instrumentation or a focused benchmark that makes per-token allocation visible.

Acceptance Criteria

• Strict token matching preserves exact rule precedence and taxonomy metadata.
• Runtime dictionary mutation rebuilds all affected indexes correctly.
• Reentrant calls and multiple filter instances cannot share stale or mutable scratch state.
• The common strict-token lookup performs no candidate Array clone, candidate Set construction, or candidate sort per token.
• Tests cover fallback rules, multiple case variants, split-token lookup characters, duplicate buckets, overlapping rules, and metadata precedence.
• check(), analyze(), censor(), and scanner paths keep compatible UTF-16 ranges and early-stop behavior.
• Same-machine benchmarks include long clean text with many unique tokens, repeated-token text, strict early match, strict late match, and filter construction.
• A unique-token negative control demonstrates that the new structure does not trade allocation reduction for a material lookup regression.

Non-Goals

• Do not change dictionary order or matching semantics.
• Do not add cross-message state.
• Do not combine strict and loose indexes in this issue.
• Do not weaken boundary or false-positive behavior.

Dependencies

Independent of #169. Coordinate numeric pattern-id representation with #168 when both changes touch compiled matcher metadata.