Slice 250.2c: Adaptive Quantization Execution Matrix (hardware-aware router + fp16 parity)

[drussell23]

Summary

Final sub-slice of the offline ML pipeline (stacked on #69511). A hardware-aware router that picks the ECAPA execution path from live host signals every call — no static model preference — with a proof the compressed path is decision-equivalent. tests/ml/ only; structural injection (no backend/unified_supervisor imports).

• Context-aware router (AdaptiveQuantizationRouter): reads Power State (AC/Battery) + Memory Pressure (structurally aligned to MemoryPressureGate.PressureLevel, lazy-probed, injectable). Rule: AC AND NORMAL → HIGH_FIDELITY; battery OR elevated → COMPRESSED. Recomputed per call (dynamic re-evaluation proven).
• fp16 quantizer: deterministic float16 round-trip + re-normalize (stand-in for fp16/ONNX-quantized encoder).
• Parity proof: per-fixture cosine drift ≈ 2.8e-8 ≪ TAU=1e-2; one threshold (0.6112) cleanly separates B(accept)/C(reject) on BOTH paths → identical Accept/Reject verdicts end-to-end through the Phase-3 BiometricExecutionMatrix.

Test Plan

• 77 passed, 3 skipped (real-ECAPA seam); tests/ml/ only
• Routing matrix (4 cases) + dynamic re-evaluation (power flip changes path)
• Cosine drift ≤ τ with huge margin; HF-vs-compressed verdict equivalence (B accept / C reject on both)

🤖 Generated with Claude Code

---

Summary by cubic

Adds a hardware-aware ECAPA router that picks the execution path per call from live power and memory pressure, plus an fp16 quantizer with tests proving the compressed path is decision-equivalent to high fidelity. Tests live in tests/ml/ only and use structural injection (no backend imports).

• New Features
  • Context-aware router: AC + normal pressure → HIGH_FIDELITY; battery or elevated → COMPRESSED. Recomputed every call with optional JARVIS_ECAPA_FORCE_PATH override. Probes are lazy, fail-soft, and structurally injected.
  • fp16 quantizer: float16 round-trip with L2 re-normalization; deterministic float32 outputs.
  • Parity: cosine drift ≪ 1e-2 with wide margin; a single threshold cleanly separates B(accept) and C(reject) on both paths via the Phase‑3 BiometricExecutionMatrix.

^{Written for commit 0056948. Summary will update on new commits.}

[github-actions]

📊 PR Size Analysis

Size: L
Files Changed: 3
Lines Added: +572
Lines Deleted: -0
Total Changes: 572

💡 This PR is large. Ensure it focuses on a single feature or fix.

Review Checklist

• All tests passing
• Code follows project style guidelines
• Documentation updated if needed
• No sensitive data exposed
• Breaking changes documented

[drussell23]

HELD (not merging this session) per operator review: same rationale as #69511 — hardware-aware quant router is production-shaped, no consumer yet, parity proven against an fp16 stand-in. Hold until the real ECAPA execution paths exist. Green + isolated; safe to land later.