perf: accelerate call-local expansion and substitution

[Qubitium]

Summary

• render exact Match.expand templates with one to eight numeric, explicit, named, or literal-backslash tokens directly from the immutable C Match snapshot
• accelerate exact default-count Pattern.sub / subn templates containing one valid numeric or named reference through a guarded C translation path
• keep all temporary parsing and output state call-local; no new persistent cache entries, unbounded collections, or cross-thread retained captures
• preserve compatibility-parser fallback for ambiguous references, unsupported escapes, non-ASCII names, subclasses, bounded substitution counts, and templates above the fixed token bound
• stabilize the API benchmark harness against CPython call-site specialization and macOS P/E-core migration artifacts
• document pinned Python 3.10 and free-threaded Python 3.14t results in the README

Performance

Pinned same-policy A/B measurements include:

• single numeric Match.expand: 16.7x on 3.14t and 20.8x on 3.10
• named Match.expand: 10.6x / 28.7x
• two-reference expansion: 9.6x / 45.3x
• three-reference expansion: 8.2x / 9.2x
• eight-reference expansion: 6.3x / 6.3x
• literal-backslash plus named expansion: 11.7x / 13.9x
• single-reference Pattern.sub: approximately 1.9–2.1x for numeric/explicit forms; named reaches 1.9x / 10.9x

Figures are listed as Python 3.14t / Python 3.10 unless shown inline otherwise.

Accuracy and safety

• differential tests cover text, bytes, Unicode, group 0, leading-zero and multi-digit references, unmatched captures, mixed numeric/named forms, duplicate PCRE names, and mutable-buffer snapshot semantics
• checked length arithmetic and fixed eight-token stack storage bound allocation math and parsing state
• shared-Match and shared-Pattern tests exercise concurrent expansion/substitution on Python 3.14t with GIL disabled
• invalid or extended syntax returns to the existing compatibility parser rather than approximating behavior

Validation

• Python 3.14t/GIL=0: 576 passed, 15 skipped, 545 subtests
• Python 3.14t Python modules: 100% statement coverage and 100% branch coverage (1,826 statements / 652 branches)
• Python 3.10: 573 passed, 18 skipped
• randomized API clobber: 60 seconds passed (seed 327942444249519898257797889077606415028)
• concurrent GIL=0 clobber: 60 seconds passed (seed 268836252769390426394361496133289920377)
• Ruff formatting/checks and git diff --check pass