nfa: +refactor Dispatch deterministic automata to scalar fast paths

[Adda0]

determinize() on an already deterministic input wrapped every state in a
one-element StateSet, reset a synchronized iterator per state, and hashed
and copied the singleton set per transition. A scalar depth-first
renumbering with the same stack discipline and increasing-symbol visit
order produces the identical output (same state numbering, finals and
edges, singleton subset_map entries, identical macrostate_discover call
order) without any subset machinery.

Nfa::is_in_lang() built a StateSet per letter through post() even when
the run is unique. While the walked symbol posts have at most one
target, scalar loop-carried reading suffices; from the first
nondeterministic post on, the subset walk continues from the state
reached with the remaining letters, keeping the answer exact without
reading the prefix again.

Nfas that only need a deterministic language representation no longer
run the determinization machinery: the new as_dfa() helper returns a
copy (or moves an rvalue) when the input is already deterministic and
determinizes otherwise, and the replace.cc marker builders use it.
complement_classical() already had the same dispatch locally.

Measured on a 5000-state, 8-symbol DFA (GCC, -O2): determinize 5.79 ms
to 3.60 ms (was 3.6x over a deep copy, now 2.3x), is_in_lang of a
4001-symbol word 3.55 ms to 0.22 ms.

union_det_complete() verified its deterministic and complete
preconditions only with MATA_ASSERT, so a Release build silently
miscomputed the union of unsupported inputs; it now throws
std::invalid_argument.

Fixes part of #801.

---

Rebased from a local stack of fixes onto current devel as a standalone commit. Timing figures above were measured on that stack and were not re-measured on devel; release build and unit tests pass on devel + this commit.

[greptile-apps]

Confidence Score: 4/5

[Medium impact] Optimizes deterministic automata dispatch with scalar fast paths.

The PR appears safe to merge, with a non-blocking improvement to avoid repeating the membership prefix.

Findings

1.  Long prefixes are read twice ▶

Summary

The PR adds scalar paths for deterministic determinize calls and unique runs in is_in_lang.

• Deterministic automata skip subset work when they are determinized.
• Single-run word checks follow one state at a time.
• Replacement markers reuse deterministic regex automata.
• Deterministic complete unions reject inputs outside their contract.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A["determinize(aut)"] --> B{"aut.is_deterministic()"}
  B -->|Yes| C["Scalar state renumbering"]
  B -->|No| D["Generic subset construction"]
  E["as_dfa(aut)"] --> F{"aut.is_deterministic()"}
  F -->|Yes| G["Copy or move unchanged"]
  F -->|No| A
  H["is_in_lang(run)"] --> I{"One initial state; no epsilon or prefix mode"}
  I -->|Yes| J["Walk single targets"]
  I -->|No| K["Generic word walk"]
  J -->|Multiple targets| K
  J -->|Word ends| L["Check final state"]

Loading

_{Reviews (1) · Last reviewed commit: "nfa: +refactor Dispatch deterministic au..." · Reviewed by Greptile}