[andbe91]

CircuitTranslationTracker.append_operation appended every operation with cirq.Circuit.append(op, strategy=cirq.InsertStrategy.INLINE), which rebuilds the current moment on each append, so converting a TICK with n operations took O(n^2) time. TICKs and REPEAT blocks were also accumulated by growing a cirq.Circuit.

Track the moment being filled as a list of operations plus the qubits, measurement keys and control keys it uses, and build each cirq.Moment once. An operation starts a new moment under the same conditions as cirq.Circuit's INLINE placement in cirq >= 1.7: a shared qubit, or a measurement key that clashes with the moment's measurement or control keys. Older cirq only checked qubits there, so with those versions, measurements that share a key (only possible with single_measure_key) are now put in separate moments, matching newer cirq. Moments are built with cirq.Moment.from_ops when available, falling back to with_operations for operations that list a qubit more than once (e.g. OBSERVABLE_INCLUDE X0 Z0), which cirq.Moment(...) rejects but the old append path accepted. Completed moments are collected in a list and turned into a circuit once at the end.

Checked against the previous implementation with cirq 1.8.0.dev: identical reprs for 5,000 random circuits x 4 option sets (default, flatten, single_measure_key, both), covering nested/tagged/count-1 REPEAT blocks, MPP, MXX/MYY/MZZ, SPP, feedback and sweep controls, noise channels, QUBIT_COORDS (with collisions), SHIFT_COORDS, tags and repeated qubits within a TICK, plus 66 conversions of generated surface, repetition and color codes. With cirq 1.6.1 the only differences are the same-key measurement splits described above.

Benchmark (best of 5):
surface code d=9, 12 rounds 48.2 ms -> 25.8 ms
surface code d=9, 12 rounds, flat 231.0 ms -> 116.7 ms
surface code d=15, 30 rounds, flat 1823.6 ms -> 911.3 ms
1000 H gates in one TICK 21.1 ms -> 6.1 ms
4000 H gates in one TICK 159.1 ms -> 24.7 ms