`state_reset()` does not clear `save_marks` after an aborted match

Bug description:

state_reset() in Modules/_sre/sre.c clears the per-match state at the start of each match in a scanner loop, including state->repeat, but not state->save_marks, which is zeroed only in state_init(). A match aborted by a signal returns straight out of SRE(match), so that counter stays above zero for the rest of the iterator and every later match saves and restores marks on backtracking that it would otherwise skip. Only finditer() and pattern.scanner() can observe it, because the findall/sub/split loops abandon the state on error. Results are unchanged, only the cost.

$ cat repro.py
import re, signal, time

class Alarm(Exception): pass

def boom(*args): raise Alarm
signal.signal(signal.SIGALRM, boom)

PAT = r"(z)*" + "(a)" * 200 + r"(?:xy|x)" * 300
SUBJECT = "z" * 400000 + ("a" * 200 + "x" * 300) * 400

def run(interrupt):
    it = re.compile(PAT).finditer(SUBJECT)
    if interrupt:
        signal.setitimer(signal.ITIMER_REAL, 0.02)
        try:
            next(it)
        except Alarm:
            pass
        signal.setitimer(signal.ITIMER_REAL, 0)
    next(it)
    t = time.perf_counter()
    spans = [m.span() for m in it]
    return time.perf_counter() - t, spans

clean, clean_spans = run(False)
dirty, dirty_spans = run(True)
print("clean iterator          : %.4fs, %d matches" % (clean, len(clean_spans)))
print("resumed after one abort : %.4fs, %d matches" % (dirty, len(dirty_spans)))
print("same results            : %s" % (clean_spans == dirty_spans,))

$ ./python repro.py
clean iterator          : 0.0087s, 399 matches
resumed after one abort : 0.0238s, 399 matches
same results            : True

Expected: an iterator resumed after an aborted match costs the same as a clean one.

CPython versions tested on:

CPython main branch

Operating systems tested on:

Linux

Linked PRs

• gh-157707

Beginne in Modules/_sre/sre.c bei state_reset und den scanner-loop-Pfaden, die von finditer() und pattern.scanner() verwendet werden. Führe zuerst den Vergleich mit repro.py aus und überprüfe anschließend, dass ein abgebrochener Match die spätere Iteration nicht verlangsamt und dass die fortgesetzten Spans unverändert bleiben.