Add verifying clobber suite: accuracy/leak/crash fuzzing on all cores - 1

[Qubitium]

Summary

Adds tests/test_clobber_verify.py — a verifying clobber suite. The existing test_clobber/test_clobber_thread hammer the API with mostly well-formed inputs and only require "no unexpected exception"; this suite generates random but valid regexes of all kinds and verifies the accuracy / leak / crash properties of every operation, on all cores − 1 workers (31 on the 32-core test box).

What it checks

Phase	What is verified
Differential accuracy (cores−1 threads)	Structured generator producing re-compatible patterns (literals, classes, greedy/lazy quantifiers, alternation, capturing/named groups, backrefs, conditionals, lookarounds, anchors, inline flags, str + bytes). Every op — search/match/fullmatch with random pos/endpos (incl. astral-plane subjects), findall, finditer, split, subn — compared span-for-span, group-for-group against re.
PCRE2-extended invariants (cores−1 threads)	Second pool adds \p{..}, atomic groups, possessive quantifiers, \R, \Q..\E over subjects up to 2 KiB; checks structural invariants: spans in-bounds & ascending, m.group(i) == subject[start:end], findall/finditer/subn counts agree, search equals first finditer match.
Shared-finditer exactly-once	One iterator drained concurrently by all workers; harvested union must equal the single-threaded reference — no lost, duplicated, or torn matches (directly exercises the FindIter lock from #108 for accuracy, not just absence of crashes).
Leak stability	Steady-state rounds over a fixed case set; asserts bounded sys.getallocatedblocks() and RSS growth.
Crash/deadlock	Shared-deadline join watchdog; a hang dumps all stacks and names the in-flight pattern of each wedged worker. On free-threaded builds the suite also asserts sys._is_gil_enabled() is False.

Generator soundness guards (each one prevents a class of false positives found while building this): group numbers tracked by opening-paren order so backrefs/conditionals only reference closed groups; nullable group bodies are never quantified (re/PCRE2 empty-repeat semantics differ); boundedness propagates into nested groups plus a per-pattern unbounded-quantifier budget, so the limit-less re oracle can't hit exponential/deep-polynomial backtracking.

Upstream PCRE2 10.46 findings (worth reporting to PCRE2Project)

The fuzzer immediately surfaced three engine-level wrong-result behaviors, all reproduced with raw libpcre2-8 via ctypes (no pypcre involved):

1. JIT start-optimization loses matches — 0?a[b ]{0,2} on b' ab ': pcre2_jit_match → no match, pcre2_match → (1, 4). Compiling with PCRE2_NO_START_OPTIMIZE makes JIT agree, implicating the JIT fast-forward scan with an optional first atom. Many variants found (e.g. (?s)(?i)[_ ]?b.{3,5}2+\w?, d?1([a-f]{0,2}(?=[^_]))-). Since pypcre enables JIT by default, affected patterns silently return wrong results through no fault of pypcre — you may want to raise this upstream and/or consider a mitigation. The suite checks JIT parity separately and reports divergences without failing (strict mode: PYPCRE_CLOBBER_STRICT_JIT=1).
2. Auto-possessification changes results around explicit possessives — (\S*?|.?2*)(b{2,2}|(c??[a-fg-k ]+|[\d\w]??_)){1,1}(\-{1,2})?+a on b'd_S2ea1-' fails to match unless compiled with PCRE2_NO_AUTO_POSSESS (re: (0, 6)). Excluded from the differential pool.
3. {0,0} zero-repetition groups — ((z)x|(?=w)y){0,0} (reduces to the empty pattern) matches nothing; re matches empty at every position. Excluded from the differential pool.

Results

• Free-threaded 3.14.7, GIL confirmed off at runtime, 31 workers, 5-minute run: 8,655,968 differential cases + 5,943,857 invariant cases — zero pypcre accuracy failures, leak growth 74 blocks / 0 B RSS, no crashes, no deadlocks.
• GIL 3.14.7: 270k + 47k cases in the 60 s configuration — clean.
• Full pytest suite green on both builds with the new module included.
• Tunables: PYPCRE_CLOBBER_VERIFY_SECONDS (default 45 s differential phase), PYPCRE_CLOBBER_SEED for reproduction (every failure message embeds the seed and full case).

🤖 Generated with Claude Code