[codex] Split parser single-branch path

[adamziel]

What changed

This stacked draft PR builds on #375 and gives the parser a dedicated fast path for the common case where FIRST-set dispatch resolves a rule to exactly one candidate branch.

• Splits single-branch parsing from the multi-branch fallback in WP_Parser::parse_recursive().
• Keeps the existing multi-branch backtracking path for ambiguous dispatch.
• Changes WP_Parser_Node::has_child() from count() > 0 to ! empty() for the hot AST completion check.

Why

After #375, branch dispatch commonly resolves directly to a single branch. The parser still paid loop/index bookkeeping designed for multiple candidates. This change duplicates a small amount of branch parsing code so the dominant single-candidate path can avoid that work while keeping the generic parser architecture intact.

This is deliberately still compact: it does not generate a large parser or expand the grammar on disk.

Performance

Benchmarks are noisy on this machine, so I compared this branch against the stacked base branch immediately before opening this PR.

• Base [codex] Speed up MySQL lexing and parsing #375 branch: 69,577 queries in 11.55021s @ 6,023 QPS
• This branch: 69,577 queries in 10.90099s @ 6,382 QPS

That is roughly a 6% incremental parser improvement on top of #375 in this run.

Parser size constraint

src/parser/*.php plus src/mysql/mysql-grammar.php remains under the requested 200 KB on-disk cap:

• 95,298 bytes total

Validation

• git diff --check
• php -l on modified parser files
• composer run test -- --filter 'WP_MySQL_(Lexer|Server_Suite_(Lexer|Parser))|WP_Parser_Node'
  • 143 tests
  • 1,421,037 assertions
• composer run test
  • 667 tests
  • 1,427,673 assertions
  • 2 skipped, 2 incomplete
• php packages/mysql-on-sqlite/tests/tools/run-parser-benchmark.php

Notes

I also tested broader parser reshapes before settling on this smaller fast path: inline terminal matching, in-memory single-fragment expansion, lightweight fragment arrays, direct branch-array dispatch, and by-reference parser cursor state. Those were slower in this codebase, so they are not included here.

[adamziel]

Closing per request; scrapping this parser rearchitecture experiment.