blueqat 2.1.0 — Exchange-Only spin qubits, correctness overhaul & cloud groundwork

This release adds first-class support for exchange-only (EO) silicon spin qubits, closes two full audit rounds of correctness fixes (including physically wrong results, not just crashes), reaches feature parity with other SDKs on everyday circuit APIs, and lays the groundwork for API-key cloud access. The default branch is now main. The test suite grew from 1,357 to 1,588 tests.

🧲 Exchange-only spin qubits (blueqat.eo)

Each logical qubit is encoded in 3 physical spins (decoherence-free subsystem); every gate is built purely from Heisenberg exchange pulses — the only native operation of semiconductor spin-qubit hardware.

• exch(theta) gate: U(θ) = exp(−iθ/2(SWAP−I)); θ = π is an exact SWAP. Runs natively on both simulation modes, autograd-friendly.
• Encoding & verification (eo.encoding): codewords for both gauge (total-Sz) sectors, leakage detection into the S=3/2 quadruplet, logical action extraction.
• Analytic pulse tables (eo.sequences): RZ = 1 pulse, X/H = 3 pulses, and the serial Fong–Wandzura CNOT (28 nearest-neighbor pulses), verified to act as an exact CNOT in all four gauge sectors including the phase (true gauge independence), after eoqrid (MIT) and Weinstein et al., Nature 615, 817 (2023).
• 'eo' transpiler backend: Circuit(2).h[0].cx[0,1].run(backend='eo') → an exchange-pulse circuit on 3n spins. Encoded Bell state verified at fidelity 1.0 with ~1e-32 leakage.
• Differentiable pulse synthesis (eo.optimizer, PyTorch autograd end to end): any logical SU(2) in 4 constant-amplitude pulses (fidelity

  1−1e-9); synthesize_2q re-calibrates drifted 2-qubit sequences — 0.05 rad perturbations on all 28 FW pulses recover an exact CNOT in ~1 s.

• Pulse schedules (eo.schedule): JSON-ready, versioned time-resolved schedules with ASAP parallel packing, round-trippable, cloud-compatible.
• quantize_sequence for discrete-duration (clock-tick) hardware constraints.

🛠 Correctness fixes (two audit rounds)

Physically wrong results (silent):

• get_time_evolution() implemented exp(+itP) instead of exp(−itP).
• Y-basis rotations in get_energy()/get_time_evolution() used RX(−π/2), flipping the sign of every term with an odd number of Y operators.

Crashes and broken utilities:

• margolus macro (NameError), draw macro (dead ibmq backend), flatten()/JSON serialization on 3-qubit gates, to_qasm() on sx/sxdg/crx/cry/crz/rxx/ryy/rzz/zz, Circuit.dagger() measurement handling, backend='composer', drawer with tensor-valued gate parameters, get_measurement_sampler's 2^24-category multinomial cap.

Silent misbehavior:

• 19 fixed gates accepted bogus parameters (e.g. x(0.5)[0], and the QASM parser's zz angle was silently dropped) — now ValueError.

✨ SDK-parity features

• Circuit.depth(), Circuit.count_ops() (slice-expanded, barrier-aware)
• Circuit.probs(qubits=None) — differentiable measurement probabilities with marginalization (PennyLane-style)
• Circuit.expect(hamiltonian) — differentiable expectation values
• iswap/iswapdg gates, barrier operation (identity in simulation, real barrier statement in QASM output)

☁️ Cloud access groundwork (blueqat.cloud)

• API-key resolution: configure() > BLUEQAT_API_KEY env > ~/.blueqat/config.json (owner-only 0600 permissions, key masked in repr)
• backend='cloud' submits the versioned JSON circuit schema; transport is injectable until the public endpoint ships.

🧪 Testing & examples

• New suites: gate identities, QFT vs DFT matrix, teleportation, dagger round-trips, statevector/tensornet cross-checks, time evolution vs expm(−itH), gate unitarity/dagger/fallback consistency, analytic-gradient checks in both modes, EO physics (48 tests).
• Examples: grover_search.py, qft.py, teleportation.py, exchange_only.py added (all self-verifying).

⚠️ Notes

• The default branch was renamed master → main.
• get_time_evolution() now follows the standard exp(−itP) convention; QAOA results are unaffected (optimizers absorb the sign), but code relying on the old inverted convention should negate its time parameter.
• Fixed gates now reject spurious parameters; previously-silent misuse like x(0.5)[0] raises ValueError.