Make adaptive compute decision-safe and runtime-parity verified

[kai9987kai]

What changed

• Added progressive accepted-probe auto compute that reuses accepted probe outputs instead of recomputing them.
• Made the post-head decision verifier the sole early-exit authority while it is active.
• Added ordered top-3 persistence, a checkpoint/workload-calibrated 0.0005 protected-margin floor, and exact three-cycle reference fallback.
• Carried the compute contract, defaults, telemetry, and diagnostics across source, browser, terminal, and packaged-runtime surfaces.
• Added bounded disagreement evidence, strict dual-mode fidelity gates, isolated packaged-response verification, numerical JSD hardening, and deterministic import-order coverage for the root compatibility entrypoint.
• Restored the route-review reconstruction visualization and documented the accepted and rejected release evidence.

Why

The first strict candidate exposed two failure classes: legacy latent-convergence exits could bypass output-decision verification, and a 0.0001 margin still allowed decision-boundary changes. In addition, focused tests could import the canonical web module while full-suite collection cached the root compatibility wrapper, which omitted one private compute-default helper.

The final design makes uncertified requests return the exact trained three-cycle ACT mixture and keeps diagnostic distribution metrics shadow-only.

Impact

• The accepted 4,096-request dual-mode gate produced zero top-1, ordered-top-3, top-3-set, or exact-output disagreements.
• Mean compute fell to 2.0378 cycles, a 32.07% reduction from the fixed three-cycle reference.
• Release-runtime weighted latency improved 4.76%; isolated-verifier weighted latency improved 7.23%.
• Progressive accepted-probe auto compute matched the legacy controller exactly across 256 requests while reducing forward evaluations 31.25% and weighted latency 30.09%.

These results are scoped to the frozen checkpoint, workload, seeds, CPU configuration, and prompt matrix; they are not a universal reasoning-quality guarantee.

Validation

• python -m pytest -q — 706 passed, 2 skipped
• strict 4,096-request dual-mode decision-fidelity gate — passed with zero disagreements
• frozen 16-prompt source/package response-fidelity gate — passed with zero mismatches
• 256-request progressive-controller gate — passed with exact output parity
• live source and packaged browser checks — defaults, blank-field fallback, verifier telemetry, and console logs passed
• python -m compileall -q source runtime_python
• python -m pip check
• python source/sync_runtime_model_variants.py --check
• python source/generate_studio_runtime_manifest.py --check
• git diff --check