Plumbing for the TDFA rewrite: Copy op, inline records, accept field, parallel moves

[hhugo]

Preparation for the tagged-DFA rewrite that fixes #199 (series: #200, #201, #202). Four commits that add everything the rewrite needs outside the compiler's determinization, without the compiler producing any of it yet. Generated code and every expect output are byte-identical (dune runtest prints no diff after each commit); the review check is that nothing new is emitted.

Commits:

1. Copy tag operation. Copies the raw contents of one memory cell into another, preserving the position/value encoding. Runtime __private__copy_mem, PPX generation, DOT label tN<-tM. Not emitted yet.
2. Inline records for tag_op. Set_position {dst}, Set_value {dst; value}, Copy {dst; src}: the rewrite manipulates these in allocation, register moves and renaming, where named fields read better than tuple positions.
3. accept : accept option, with accept = { rule; final_ops }, replaces finals : bool array on DFA states. Rule priority is resolved in the compiler (accepting_rule) instead of the PPX's best_final, and the accepting state carries the operations to run on entry just before Sedlexing.mark. The rewrite will use them to copy the accepting path's working registers into the canonical cells read by binding extraction; the list is [] for now, so gen_tag_ops wraps nothing. Nothing driving the automaton needs a state's other accepting rules, so finals goes away and the DOT printer labels an accepting state with the winning rule only; no expect output listed more than one rule, so the graphs are unchanged.
4. Parallel-move gen_tag_ops. A transition's operations form a parallel move: a Copy reads its source as it was before the list ran. Sources the list also writes are saved in let-bound locals through the new raw accessors __private__mem_get / __private__mem_set; op_dest is exported for it. No Copy is emitted yet.

With these in, the rewrite's diff outside src/compiler/sedlex.ml is confined to expect outputs (plus the matching one-line changes in the #202 interpreter).

🤖 Generated with Claude Code

[pmetzger]

Very cool!