Add production tuner defaults and generic AI Arena language edges

[Valessc2]

Scope

Post-32/32 V1 polish that turns the Tuner's engineering priors and generic multi-language consumption model into explicit code and CI.

Production seed policies

• one Defaults.hpp authority
• 20 operational samples
• 50-sample certification seed
• 2% + 1 microsecond default promotion margin
• z = 1.96 champion confidence seed
• 1% + 50 microsecond tuner/control-overhead ceiling
• workload seeds 8..1024, batch seeds 16..1024, spatial-cell seeds 2/3/4/6/8/12
• hardened TunableRegistry rejects non-finite and malformed Integer/Choice contracts

These are consumer-overridable priors, not hard bot-performance ceilings.

No bot performance cap

The compute-budget governor is OFF by default. SC2AIOptimize cannot shed work unless the consumer explicitly delegates a compute envelope. The helper then derives 55/70/95% resume/soft/hard hysteresis; a 40 ms consumer-owned envelope maps to 22/28/38 ms.

Generic AI Arena language edges

• direct generic C++ remains the cheapest native surface
• existing optional Python buffer adapter remains independent
• optional stable shared C ABI (sc2opt_c) with caller-owned buffers
• thin Rust, Go and .NET wrappers over the C ABI
• Java 17 JNI edge using direct ByteBuffers
• Node.js N-API edge using typed arrays
• no MMEvo-specific or framework-specific adapter is introduced

The C ABI and all non-C++ language edges are OFF by default and are not linked by sc2opt::sc2opt, so a direct C++ consumer such as MMEvo pays none of their cost.

CI expansion

• existing Ubuntu/Windows core lanes remain unchanged with all new edges OFF
• existing Ubuntu/Windows Python lanes remain
• new Ubuntu/Windows C ABI build + runtime tests
• new Ubuntu language-edge compile lane for Java 17, Node N-API, Rust, Go and .NET
• existing sanitizers, clang-tidy, installed-consumer/LTO and AVX2 gates remain

Handoff intent

Once all gates are green and this PR lands, MMEvo can consume current main through the direct generic C++ surface and begin representative calibration. Real bot/runtime speedup remains evidence to be measured, not pre-claimed.