Compare the oracle reference matcher with ocamllex

[hhugo]

Stacked on #202 (the GitHub base is a copy of its branch): ocamllex as a second reference for the oracle harness. Tests only.

• Vendored ocamllex (test/ocamllex_vendored, from OCaml 5.4.0, sync.sh to refresh): Lexgen.make_dfa and the modules it needs, unchanged. These files are LGPL 2.1 with the OCaml linking exception; they are only linked into the test runner.
• Bridge (test/ocamllex_oracle/ocamllex_oracle.ml): a rule set over Ir.t is translated to ocamllex's syntax (alternatives in order, repetition written iterate-first), compiled with Lexgen.make_dfa, and the automaton is run on the input the way ocamllex's engine does. The result is compared with the reference matcher of Add an oracle test harness: reference matcher vs compiled tagged DFA #202.
• Harness change (test/oracle/oracle.ml): sweep takes the check to run and the printer for failures, and the generators take ~eof so a comparison that cannot model end of input can leave it out of the patterns.

What the comparison found:

• On the CI-sized sweeps (2000 single-rule and 1000 two-rule cases, no eof) and on 300k local cases, ocamllex and the reference agree on rule and length everywhere.
• Captures differ in one way only: ocamllex gives the earlier sub-pattern the longest match where sedlex takes the left alternative (('c' | 'd' ['a'-'b']) (['b'-'d']* as y) on db: sedlex y="b", ocamllex y=""). The comparison accepts ocamllex's bindings when they are another parse of the same lexeme and prints them after the reference's; a hand-written case shows two such pairs.
• ocamllex 5.4.0 itself miscomputes a capture preceded by an overlapping loop, the same shape as sedlex's Rewrite determinization as a Laurikari TDFA for correct capture semantics #199: ('a' | ['b'-'d']*) (('c' | "bb") as y) (['a'-'d'] ['a'-'d'])* (['a'-'b'] as w) on ccdb gives y="", which no parse admits (y="c" is the only one). Reproduced with the real ocamllex binary, not just the vendored simulation. Pinned as an ERROR line in a hand-written case.

End of input is not translated: ocamllex can read it repeatedly, sedlex reads it once.

Commits (on top of #202):

1. Vendor ocamllex's DFA compiler from OCaml 5.4.0.
2. Let the oracle's sweep and generators serve a second comparison.
3. Compare the reference matcher with ocamllex.

🤖 Generated with Claude Code

[pmetzger]

@hhugo You should have the ability to commit such things on your own at this point...

[pmetzger]

@hhugo Do you intend to commit this?

[hhugo]

@hhugo Do you intend to commit this?

I need more time to work on this. The current implementation on master is not working properly. I've a fix locally, but still need to work on the specification of the behavior and work on testing (such as this PR).