Implement corrected 0.40 performance tranche

[rbroderi]

Summary

• merge the 0.39 sparse-set performance work
• remove the three 0.40 whole-algorithm matrix/tree compilers
• retain and generalize GC, accounting, scalar-entry, record, typed-IR, and sparse-set improvements
• add admission telemetry, a permanent generality policy, metamorphic tests, and a three-program A/B gate
• add the expanded language benchmark corpus

Validation

• 653 Python tests pass on Python 3.14
• focused corrective and language tests pass on Python 3.13
• all 24 unchanged official Lua 5.5.1 probes pass
• git diff --check passes