Add lazy optimized MRV indexing

[xtraid]

Summary

• add a lazy private packed MRV bucket index only to the optimized serial solver
• preserve the reference linear scan, row-major tie break, solver semantics, trace events, and ownership
• expose direct MRV probe/storage metrics and document the isolated mechanism

Evidence

• switch-off/switch-on comparison across all 25 native benchmark cases preserved status and every non-MRV deterministic metric
• generic unconstrained SAT reduced MRV cell scans from 43,897,478 to 18,274 and median solve time from 149.782 ms to 5.792 ms (-96.13%)
• largest observed median control regression was +2.24%, within the declared guardrail; no Yang-Zhang speedup claim is made

Validation

• make check: 17 C test binaries, 151 Python tests, Pages, native and cross-engine smoke
• 262 renderer tests and isolated pdfLaTeX dossier smoke
• strict GCC/Clang, ASan/UBSan/LSan, GCC analyzer, Memcheck, complete Cachegrind
• informational C/Python coverage and deterministic 2,000-run parser fuzz smoke
• diff, modes, credential patterns, and versionable artifacts clean

Scope

No TaskPlan, OpenMP, structural scheduling, trail compaction, solver reference rewrite, or renderer change is included.