Prototype table-driven tokenizer bulk scanning

[isomorphisms]

Refs #38.

This adds a deliberately narrow, test-only byte tokenizer experiment under the existing Android-prepaint JVM harness.

What it compares:

• a scalar oracle that makes the lexical decision one byte at a time;
• a candidate with a small byte-class table plus state × class -> action dispatch;
• a bulk path that consumes maximal APPEND_RUN spans before returning to exceptional-byte dispatch.

The event contract is pinned explicitly and the candidate must match it across every single two-chunk split and one-byte chunking. Fixtures cover quoted attributes containing </>, entities, malformed entities, NUL, unclosed quotes, and raw UTF-8 handoff including splits inside multibyte sequences.

The long-run test also proves the candidate is actually taking the bulk path: 4096 ordinary bytes are consumed as bulk work while exceptional dispatch stays below 32 operations.

Deliberate boundaries:

• this is a lexical experiment, not an HTML conformance tokenizer;
• it does not integrate with renderer/prepaint production code yet;
• UTF-8 is handed off losslessly as raw bytes rather than decoded here;
• SIMD/word-at-a-time scanning is intentionally absent. It should be a later implementation comparison against the same oracle, not a prerequisite for validating the architecture.