Add cheap generic Tuner through S24

[Valessc2]

Scope

Completes the S20-S24 Tuner package on top of current S19 main without moving bot semantics or framework types into SC2AIOptimize.

S20 — net benefit

Existing baseline/OFF gate remains candidate zero and continues to require correctness-qualified total-cost evidence.

S21 — TunableRegistry

Adds fail-closed registry validation plus ID/name lookup. Invalid specs and duplicate IDs/names are rejected. Consumer code still owns meaning, storage and safety.

S22 — champion/challenger

Adds allocation-free Welford running statistics, current-champion vs challenger evaluation, explicit BaselineOnly mode, deterministic reset/evidence reset and conservative confidence separation. Invalid policies/evidence fail closed.

S23 — context/crossover

Adds compact numeric workload/hardware/state context lookup and ordered workload-size crossover bands. Unknown/invalid contexts or bands fall back to baseline by default.

S24 — persistence/provenance

Adds a strict versioned standard-library persistence record carrying source revision, SC2 build/data version, compiler, hardware and workload profile identity. Persistence is control-plane/I/O work and deliberately not part of the hot path.

Cheapness

• no ML/runtime dependency
• no hidden allocations in champion/context selection
• no sample-history retention; Welford keeps count/mean/M2 only
• no benchmarking/search inside hot kernels
• steady state is context/crossover -> already-resolved candidate ID -> execute
• baseline/OFF remains selectable

Validation

The exact S20-S24 implementation was built locally under GCC 14.2 with -Wall -Wextra -Wpedantic -Werror; the standalone tuner test suite passed 1/1. Full repository Linux/Windows CI on this PR is the cross-platform build/test proof.

Not claimed

This does not claim target-hardware speedups, calibrated crossover values, bot integration, ladder evidence or V1 certification. Those remain S25-S32 work.