[inureyes]

Summary

Every pipelined decode loop in CxxGenerator sampled step t+1 before pushing token t into token_history, so with a history-reading sampler (repetition, frequency or presence penalty, DRY) each token was sampled against a history missing the one just emitted, and the first decode token saw the prompt alone (#2090). mlx-lm's generate_step and llama.cpp's common_sampler_accept both include it.

What changed

• sample_next_step (in generate.rs) does the ordering once for all four loops (generate_streaming, generate_streaming_with_embeddings, generate_with_stats_and_embeddings, generate_with_stats). On the history path it submits the step t+1 forward with async_eval, reads token t, pushes it, and only then builds the sample, returning the token so the loop neither reads nor pushes it twice. The history-free path keeps sampling from the lazy graph in the old order.
• Server batch scheduler: unaffected. lookahead_params admits only rows that pass config_supports_fused_batch_except_bias, which rejects needs_token_history, so penalty rows never pipeline; the per-row paths read and push each token in the tick that sampled it (decode_tick.rs), and prefill pushes the first token before the first decode tick (prefill.rs).
• Tests: test_support::induction now holds InductionModel, lcg_tokens and sequential_reference for both test files. generate_history_tests runs all four entry points against the sequential reference under a windowed repetition penalty, full-history counts (the incremental SamplerState path), frequency/presence, DRY, and no penalty; the four penalty cases fail on the old ordering. Prompt lookup's penalty test checks CxxGenerator against the reference again.

mlxcel generate, run and chat output under penalties changes, intentionally.

Validation

• generate::history_tests and speculative::prompt_lookup: 44 passed; reverting the ordering fails 4 of the 5 history tests and prompt lookup's penalty test.
• Real checkpoints on GB10: before/after release binaries built from one tree (only generate.rs swapped), greedy, median of 3, interleaved (load median 2.9):

Case	no-penalty after/before	no-penalty output	--repetition-penalty 1.3 --repeat-last-n 3 after/before	penalty output
Qwen3-1.7B write	1.004x	identical	0.998x	identical
Qwen3-1.7B story (800)	0.990x	see below	1.005x	differs
Qwen3-8B write	1.007x	identical	1.010x	identical
Qwen3-8B summary	1.003x	identical	1.000x	identical
Llama-3.1-8B story (800)	1.005x	identical	0.997x	differs

Throughput is unchanged on both paths. Penalty output changes only where a stale window would have chosen differently (the Llama story's before-penalty reply has a doubled word the after reply does not). The Qwen3-1.7B story differs without a penalty because plain greedy decoding of that reply is not reproducible run to run on this host: three runs of the unchanged binary diverge from each other at characters 1300 to 1351, the same range as the before/after divergence (1432). That is pre-existing and outside this PR.

• Reviews: pr-reviewer and pr-security-checker found no CRITICAL, HIGH or MEDIUM issue in the change; their LOW notes on diagnostics are now documented at sample_next_step. A pre-existing gap they noted (mirostat and adaptive-p state is not carried by CxxGenerator) is unreachable today: generate and chat fix both off, and the server never uses CxxGenerator.

Closes #2090

[inureyes]

Merging with WebUI installed artifact red, by maintainer decision.

• The job failed five times on 19222848, every time at its pre-check (GPU compute processes are already present before Activity gate), including one rerun on an idle host (load 1.3). The check prints a count only; nvidia-smi on the host showed no compute process before or between the reruns, and the steps before the gate produced the same output as main's run of 680f64ee, which got past the pre-check and failed later on its activity-performance stage.
• This PR cannot change what that job runs: it touches CxxGenerator's decode loops (generate, run, chat), and the server path the job exercises goes through BatchScheduler and never constructs a CxxGenerator.
• The other 15 checks pass. Locally on GB10: 80 tests in generate::history_tests, speculative::prompt_lookup and generate::tests, clippy -D warnings, and the before/after release-binary comparison in the PR body.

The pre-check would be easier to diagnose if it printed the PIDs and process names it found.