Apply two-qubit gate noise to every pair of a batched command

[ciaranra]

Closes #791. Closes #792.

Problem

A two-qubit gate command in a ByteMessage can carry several pairs in one flattened qubits slice. Three noise models treated such a command as if it held one pair:

• DepolarizingNoiseModel::apply_tq_faults and BiasedDepolarizingNoiseModel::apply_tq_faults drew one fault and applied it to qubits[0] and qubits[1] only, so every later pair ran noiselessly (Depolarizing and biased-depolarizing noise fault only the first pair of a batched two-qubit command #791).
• GeneralNoiseModel::apply_tq_faults iterated pairs but evaluated has_leakage and the seepage loop over the whole command, so one leaked qubit removed every pair, healthy ones included (GeneralNoiseModel drops every pair of a batched two-qubit command when any one qubit is leaked #792).

The workspace sweep found the same shape in exp/pecos-eeg/src/noise.rs (UniformNoise::noise_after_gate injected noise for the first pair only) and in two pecos-neo channels (TwoQubitChannel::handle_after_gate, PerGatePauliChannel::apply_after_gate).

Change

Every site now iterates qubits.as_chunks::<2>() and draws, faults, and judges leakage per pair. A single-pair command is byte-for-byte unchanged, including RNG consumption; the existing pinned-byte tests are untouched and still pass. CCX keeps its current behaviour (one draw on its first two qubits), pinned by a new test rather than silently changed.

Not in this PR

Splitting batched commands per pair inside the pecos-neo runner was implemented and then reverted: the independent review reproduced that it changes what crosstalk and correlated channels observe. Per-pair SkipGate suppression in neo needs a batch-preserving design and is tracked as #807.

Tests

• Batched vs consecutive single-pair commands from identically seeded models must produce identical output at p2 = 0.5 over 32 seeds, followed by a further command to detect RNG-stream drift (engines and neo).
• General noise: one leaked qubit in CX[(0,1),(2,3)] suppresses only (0,1); (2,3) survives and receives its own noise; seepage is pair-local.
• EEG: batched injections equal the concatenation of per-pair injections.
• Mutation-checked: sharing the occurrence draw across pairs, reusing the first pair's Pauli, and restoring the whole-command leakage check each make the new tests fail.

Verification

cargo test -p pecos-engines -p pecos-neo -p pecos-eeg --no-fail-fast (1852 passed), cargo clippy --locked ... --all-targets -- -D warnings and cargo fmt --check on the three crates, all clean.

How this was produced

Implemented by OpenAI gpt-6-astra (Codex CLI 0.154.0) from a task packet written by Claude Fable 5.1 in Claude Code; Claude reviewed every hunk, re-ran all verification itself, and sent findings back for fix rounds. An independent Codex review arm with no access to the implementation session also reviewed the diff. Posted at the maintainer's request.