Quantum-classical co-design loop (BL-33)

scpn_quantum_control.codesign composes the existing phase-objective, gradient-planning, open-system, control-adapter, and observer products into a deterministic local loop. It is a software-in-the-loop research surface. It does not submit provider jobs, actuate hardware, replace a plant controller, or claim operational plasma stability.

Execution sequence

LoopStepInput
  -> PhaseObjectiveSimulator (BL-47 policy + phase objective + BL-09 plan)
  -> ExponentialOrderEstimator
  -> GradientFeedbackController proposal
  -> LatencyPolicy
  -> SafetyEnvelope (+ optional BL-69 interlock)
  -> LoopStepOutput / ReplayTrace

Every record uses quantum_classical_codesign.v1. Inputs and outputs are immutable. Logical timestamps make delayed, missing, and stale-gradient cases replayable without pretending that workstation timing is a hardware deadline.

Bounded modes

Mode	Estimator input	Intended use
classical_to_quantum	Caller-supplied classical measurement	Evaluate an existing phase objective and propose a correction.
quantum_to_classical	Local simulator order parameter	Feed bounded simulator telemetry into the classical estimator.
hybrid_replay	Recorded measurement	Reproduce a complete deterministic controller trajectory.

The evaluator builds the existing analytic Kuramoto order-parameter target objective. Its exact objective gradient is the applied simulator reference. The attached BL-09 explanation is a backend-capability decision, not a claim that the analytic classical term was silently converted to parameter shift.

Policy and safety contract

• A concrete ClosedLoopExecutionPolicy is mandatory.
• Even an otherwise valid hardware policy is refused by the BL-33 evaluator.
• Missing and stale gradients select an explicit hold or abort; no previous gradient is silently reused.
• Gradient norm, update norm, and absolute parameter envelopes are checked before proposed parameters become the applied output.
• BL-69 hold and abort decisions take precedence over controller proposals.
• BL-68 active-sensing and BL-70 geometry records remain telemetry only.
• No network I/O or provider SDK call occurs in this package.

Existing-product adapters

The package consumes, rather than duplicates, the BL-67 ports:

• consume_realtime_feedback_port(...) uses RealtimeFeedbackPort;
• consume_qaoa_mpc_port(...) uses QaoaMpcPort;
• consume_cosimulation_port(...) uses the policy-gated existing partition;
• observer_inputs_from_products(...) maps BL-68, BL-69, BL-70, and BL-80 records. BL-80 contributes only decrease/hold telemetry after BL-47 approves its complete dry-run shot plan.
• adaptive_fim_proposal_port(...) maps one BL-80 step to an unapplied scalar ControllerProposal; the co-design safety envelope must still decide whether any later simulator action is allowed.

This is the versioned FUSION/CONTROL bolt-on contract. A partner repository may implement the typed port or translate its plant telemetry into LoopStepInput; it must retain ownership of plant models, realtime scheduling, actuation, and operational safety. No cross-repository import or circular dependency is required.

Optional open-system objective

OpenSystemObjectiveConfig augments the phase objective with one existing BL-16 bounded Lindblad or seeded-MCWF case at the current parameter vector. The configured weight is explicit. Dimension drift fails closed. This is local solver evidence, not measured hardware decoherence or an adjoint-Lindblad gradient claim.

Deterministic replay

from scpn_quantum_control.codesign import (
    ReplayTrace,
    build_demo_loop,
    demo_inputs,
    record_replay_trace,
    verify_replay_trace,
)

trace, outputs = record_replay_trace(build_demo_loop(), demo_inputs())
encoded = trace.to_json()
parsed = ReplayTrace.from_json(encoded)  # verifies the SHA-256 envelope
verify_replay_trace(build_demo_loop(), parsed)  # bit-identical trajectory

Replay covers inputs, optional observers, complete output records, schema, claim boundary, and digest. A changed measurement, mode, policy outcome, or controller trajectory is rejected.

Evidence and performance boundary

Run the bounded evidence writer:

python scripts/run_codesign_loop_evidence.py \
  --output data/differentiable_phase_qnode/codesign_loop_evidence.json

The committed artefact records local loop throughput, Python/platform context, the replay digest, and functional_non_isolated classification. It is not an isolated timing claim and must not be used as a PCS, provider, or hardware latency result.

The new package owns orchestration, serialisation, and safety decisions rather than a numerical hot kernel. Phase objectives, open-system solvers, realtime feedback, QAOA-MPC, and co-simulation remain with their existing Python/Rust owners. A second Rust implementation here would duplicate ownership without a measured hot-path need.

Plasma-relevance templates

plasma_objective_templates() exposes only non-operational research framing: synchronisation-target recovery and stale-observer hold. Partner validation is required. No template names an actuator command, plant operating envelope, or live control-system guarantee.

Claim boundary

deterministic local simulator orchestration over existing phase and control surfaces; no live QPU, provider submission, operational plasma control, realtime hardware, stability guarantee, or quantum-advantage claim

Authored by Anulum Fortis & Arcane Sapience (protoscience@anulum.li)