Add an oracle test harness: reference matcher vs compiled tagged DFA

[hhugo]

Tests only: a differential test harness for the compiler, in a new test/oracle library that uses sedlex.compiler directly and involves no PPX.

Two independent implementations of sedlex's matching semantics run on the same (rules, input) pairs and their results are compared; a disagreement is a compiler bug.

• Reference matcher (ref_match in test/oracle/oracle.ml): a brute-force backtracking matcher over Ir.t. It enumerates every parse in leftmost-greedy priority order (alternation prefers the left branch, Star/Plus/Rep prefer one more iteration) and keeps the first parse among those of maximal length; across rules the longest lexeme wins and the lowest-numbered rule breaks ties. End of input is a marker after the input, read at most once and adding no length (Read end of input once, and let priority decide its ties #211). This is the specification of capture semantics and the part to review. It is deliberately naive so that it is correct by inspection.
• DFA interpreter (dfa_match): runs the output of Sedlex.compile_ir the way PPX-generated code does: memory cells start at -1, a transition's operations record the position before the code point that selected it (TDFA(1): delay the tag writes of a closure to the outgoing transitions #209), the final operations run on accept before the mark (Stop copying the capture cells in Sedlexing.mark and backtrack #208), end of input is fed once as the code point -1 without advancing, and on a dead end the last mark wins. Bindings are extracted as gen_binding_code does, discriminator cells included.

Entry points: oracle prints one line per input, prefixed with ERROR and both results on disagreement; qcheck is a deterministic random sweep (fixed seed, no shrinking) over sequences of up to four elements on the alphabet a..d plus eof, with at least one capture, plus or-pattern rules binding the same name in both branches, on inputs of length 0 to 5. Failing cases print through oracle with the pattern, so an empty expect block means full agreement.

The hand-written cases cover the basic capture shapes, or-patterns, greedy disambiguation, the loop-before-capture patterns of #199 and end of input (capture before eof, eof adding no length so rule priority decides, a tie within a rule, Plus eof / Star ('a' | eof) / eof, eof). Every expect block is clean on master; the harness is what was used to check #204, #207, #208, #209, #210 and #211 as they were written (millions of random cases locally, the sweeps here are sized for CI: about 1.5 s).

The interpreter is only as faithful as its imitation of the PPX glue, so a bug in that glue rather than in the compiler can escape it.

Commits:

1. qcheck-core declared as a with-test dependency in dune-project and sedlex.opam.
2. The harness (oracle.ml), the expect tests (test_oracle.ml), and the dune library, gated on OCaml >= 4.14 like the other expect tests.

🤖 Generated with Claude Code