Add cheap runtime optimisation through S27

[Valessc2]

Scope

Completes S25-S27 on top of the merged S20-S24 Tuner without adding a fourth runtime pillar.

S25 — profiling / telemetry

• fixed-storage 65-bucket log2 latency histogram
• allocation-free observation with saturating counters
• p50/p95/p99 conservative bucket upper bounds
• compile-time-disabled scoped timer with no clock read or histogram mutation

S26 — compute-budget governor

• generic Normal/Soft/Hard latency pressure state
• hysteresis using previous completed duration
• consumer-owned Critical/Important/Optional work classification
• invalid threshold ordering fails closed to Hard
• default 22/28/38 ms seed values generalised from MMEvo Z7, without importing MME policy

S27 — auto-calibration

• caller-owned workload points and evidence
• requires baseline evidence at each point
• delegates winner selection to the single S20 ChooseNetBenefitChampion authority
• compresses adjacent winners into tiny crossover bands
• explicit measured control-overhead gate so the optimiser can reject its own bookkeeping if too expensive

VC hardening found during the slice

The existing S20 net-benefit policy accepted non-finite policy thresholds implicitly. This PR makes invalid/NaN policy values fail closed to baseline with an explicit PolicyInvalid reason.

Cheapness

• C++20/STL only; no new dependency
• no runtime allocation in histogram observation, governor updates or crossover lookup
• no benchmarking/search in hot kernels
• compile-time instrumentation-off path has no timer state/clock read
• auto-calibration is control-plane; its output is a compact candidate-ID table
• baseline/OFF remains valid for every workload range

Validation

The exact new telemetry/governor/calibration code was compiled locally with GCC 14.2 under -Wall -Wextra -Wpedantic -Werror; the runtime optimisation tests passed. Full repository Linux/Windows core + Python-adapter CI on this PR is the cross-platform gate.

Not claimed

No target-hardware crossover values, runtime speedup, bot integration or V1 certification is claimed. S28-S32 remain the proof/output/hardening programme.