[ruslan-splynx]

CPU FP64/FP32 and Metal CLI state-vector runs previously used raw execution even when benchmarks/Python exercised fusion. CLI now selects a backend-appropriate optimized executor independently of formatting; --raw retains the reference driver and --execution-stats exposes actual retained IR kinds/planning cost.

Closes #551. The previously deferred Mac validation is now complete (2026-11-01), after integrating main 4e72028. Integration preserves noise-budget/CUDA execution APIs and benchmark diagnostics. Review and green CI are required on this integrated head.

Final-state pipelines omit measurement-seeded DCE so preparation on unmeasured qubits, collapsed-final-state sampling, amplitudes and expectations survive partial measurements. Relabelling is restored before views. CPU retains small precomputed diagonal gates; a phase-polynomial pilot made FP64 QFT 54% slower despite a smaller IR plan. Metal uses its supported GPU final-state wrapper. FP32 supports dense/diagonal fusion and correct tiled replay. Stabilizer/MPS/noise/CUDA drivers and the legacy bench subcommand are unchanged; the original Rust run_circuit API remains available.

Actual public CPU CLI A/B on i5-13500 Linux, release, RAYON_NUM_THREADS=1, 32 shots/seed 7, three interleaved fresh processes per mode:

Precision / workload	Raw → optimized caller wall time
FP64 Bell n=2	1.051 → 1.095 ms
FP64 QFT n=20	929.441 → 955.058 ms (no gain; 2.8% slower)
FP64 brickwall n=20 d=12	774.322 → 537.925 ms
FP32 Bell n=2	1.146 → 1.182 ms
FP32 QFT n=20	2085.930 → 1050.777 ms
FP32 brickwall n=20 d=12	1130.771 → 579.510 ms

Public-process Criterion (10 samples, 1 s warmup, 2 s requested measurement, committed n=15 fixtures): FP64 brickwall 25.605 → 22.854 ms; FP32 QFT 37.224 → 19.194 ms; FP32 brickwall 40.369 → 24.355 ms. FP64 QFT median confidence intervals overlap, so no speedup is claimed. Tiny inputs expose added planning overhead. Comparisons use the same binary's raw/optimized modes, not historical binaries or kernel-only timers.

Fusion can change FP rounding/alias bins and hence a seed's exact histogram. Each mode is reproducible. Full n=20 brickwall verification found optimized FP32 max amplitude error 2.85e-9 and probability L1 error 5.52e-7 versus raw FP64 (raw FP32 L1 1.69e-6). CLI seed help documents the mode-dependent mapping.

Validation: 35 CLI unit tests, 40 existing command tests, GPU-init regression, three new public optimizer tests; final-state amplitude oracles for CPU FP64/FP32 including relabelling, external controls, partial measurements, phases/barriers and 64 generated circuits; existing optimized-driver oracles; three public CPU Python/CLI equivalence fixtures. Real Metal CLI raw/optimized oracle executed successfully. CUDA CLI compilation, workspace/Metal all-target clippy, formatting and diff checks pass. Raw data, Criterion samples, planning counts/costs, limitations and pending work: docs/perf/p8-01-optimized-cli.md.

Deferred Metal validation completed (2026-11-01)

Apple M5 MacBook Air, 10 CPU/10 GPU cores, 32 GiB unified memory. Release public CLI, FP32 Metal, same binary raw vs optimized, 32 shots/seed 7, RAYON_NUM_THREADS=1. One untimed process per mode, then four ABBA blocks (8 observations per mode), with diagnostics and host telemetry. Whole-process timings include planning/startup/output. No programs/services stopped; shared desktop, CPU idle median 74.68% (range 57.11–94.36%), GPU/thermals not isolated.

Metal workload	Raw → optimized median	Ratio
QFT n=20	91.931 → 42.257 ms	2.18×
Brickwall n=20 d=12	69.918 → 38.411 ms	1.82×
QFT n=24	3139.286 → 331.996 ms	9.46×
Brickwall n=24 d=12	1628.686 → 294.843 ms	5.52×

Bell n=2 has no consistent gain: 28.861 → 27.552 ms in the first series; 32.747 → 32.793 ms in the second. Public-process Criterion without diagnostics (10 samples, 1 s warmup/2 s requested measurement) gives median estimates: Bell n=2 26.699 → 27.358 ms; QFT n=15 33.196 → 28.375 ms; brickwall n=15 d=20 33.298 → 34.138 ms. Small Bell/brickwall remain no-gain cases; large-n gains are not universal.

Integrated validation: 81 release CLI tests with actual Metal explicitly enabled; 5 final-state oracle tests (1 existing ignored offline case); 4 optimized-driver oracle tests; 4 Metal oracle tests plus 1 property test. Three focused public optimizer tests passed again with Metal structural assertions. New n=16 real-Metal oracle verifies all 65,536 amplitudes against raw CPU FP64 after permutation restoration and partial measurement/unmeasured preparation. Workspace all-target Clippy, Metal CLI/backend all-target Clippy, formatting, Python syntax and changed documentation/checksums passed. Manual integration review found no blocking findings.

Raw samples, confidence intervals, logs, host telemetry, source/binary fingerprints and commands: docs/perf/data/p8-01-metal/. Report: docs/perf/p8-01-optimized-cli.md. Same-mode seed reproducibility passed; cross-mode histogram equality is not required because floating-point fusion can alter alias bins.