Work around PCRE2 10.46/10.47 JIT start-optimization losing matches

[Qubitium]

Summary

Works around the PCRE2 10.46/10.47 JIT start-optimization bug found by the verifying clobber suite (#110): the JIT fast-forward scan skips valid start positions for patterns beginning with an optional atom + literal + bounded class — pcre2_jit_match finds nothing for 0?a[b ]{0,2} on " ab " while pcre2_match returns (1, 4). Since pypcre enables JIT by default, affected patterns silently returned wrong results.

Affected-version matrix (verified with raw libpcre2-8 builds via ctypes, pypcre not involved):

PCRE2	JIT vs interpreter on the repro family
10.46 (system)	diverges (JIT loses matches)
10.47 release (built from tarball)	still diverges — all repros
current main (10.48-dev, built with its pinned sljit)	fixed

So every currently shipping release is affected; only unreleased main is fixed.

The fix

Follows the existing jit_anchor_fixup_needed() precedent exactly:

• One-time runtime probe (jit_start_optimize_broken(), run eagerly at module init): compile 0?a[b ]{0,2}, run pcre2_jit_match vs pcre2_match(PCRE2_NO_JIT) on the same compiled code against " ab ", compare results. _jit_start_optimize_broken() is exposed for introspection/tests.
• Targeted mitigation: on a probe-positive runtime only, Pattern_create injects PCRE2_NO_START_OPTIMIZE into the compile options of JIT-bound patterns. The option bit exists on every PCRE2 version (unlike pcre2_set_optimize, added in 10.43), so there is no new symbol dependency and old runtimes still import.
• No flag pollution: the injected bit is masked back out of Pattern.flags (it shares the value 0x10000 with Flag.SUBSTITUTE_MATCHED on the Python side); an explicitly caller-requested NO_START_OPTIMIZE is preserved as before.
• Fixed runtimes probe clean and compile exactly as before — zero cost, no behavior change.

Perf note: on broken runtimes, JIT-bound patterns lose PCRE2's first-code-unit fast-forward (pypcre's own literal prescan is also disabled for them since PCRE2_INFO_FIRSTCODEUNIT is no longer populated). Correctness over speed, and only where the runtime is broken.

Validation

• Empirical completeness: the clobber suite (Add verifying clobber suite: accuracy/leak/crash fuzzing on all cores - 1 #110) re-run on the affected 10.46 runtime with PYPCRE_CLOBBER_STRICT_JIT=1 (any JIT-vs-interpreter divergence fails the test): 3,549,787 differential cases, zero divergences — previously ~10 divergences per minute. The workaround covers the whole observed bug family, not just the probe pattern.
• New tests/test_jit_start_optimize_workaround.py: pins six known-diverging pattern/subject pairs against re under default (JIT) compilation — these assertions hold on both broken runtimes (via the workaround) and fixed runtimes (natively) — and asserts the flags contract.
• Full pytest suite green on CPython 3.14.7 free-threaded (GIL=0 at runtime) and 3.14.7 GIL builds against system PCRE2 10.46.

Worth reporting upstream to PCRE2Project alongside the two still-unfixed engine behaviors documented in #110 (auto-possessification changing results around explicit possessives; {0,0} zero-repetition alternation groups matching nothing) — bug 1 needs a backport request since the fix only exists on main.

🤖 Generated with Claude Code