[inureyes]

Summary

Completes the b10621 sampling semantics that issue #1436 owns in the compatibility manifest: the repeat_last_n penalty window finally reaches the sampler, the DRY sentinels and off-by-one tier are reconciled with upstream, the top-n-sigma / XTC / ignore-eos / reverse-prompt server and native surfaces land on both binaries, and the sampler-order flags accept exactly the fixed b10621 default chain. 26 manifest entries flip to supported, 3 to not_applicable with tested rejection, and the 31 entries this PR cannot honestly claim repoint to the new remainder issue #1485, added as a shard co-owner in pin.json.

Behavior changes to know about

• Server-side generations that enable a repetition / frequency / presence penalty and omit repetition_context_size (alias repeat_last_n) are now bounded to the server-wide --repeat-last-n default of 64 recent tokens, matching b10621; before this PR the flag was parsed and echoed by /props but every penalty scanned the whole history. The CLI keeps its pre-existing full-history behavior via a -1 default on the new --repeat-last-n flag.
• dry_penalty_last_n: 0 now disables DRY (upstream's sentinel); the pre-existing "0 = whole history" form moved to the explicit DRY_FULL_HISTORY, spelled -1 on the CLI flags. The DRY comparison changed from > to >=, adding upstream's base^0 penalty tier at exactly dry_allowed_length. A dry_base below 1.0 falls back to the server default on HTTP and keeps the CLI default silently, both matching b10621.
• penalty_last_n == 0 makes the penalty stage inert regardless of penalty values, so such configs stay on the fused batch path (needs_token_history gates on the windows now).

Rust versus C++ placement

No C++ or bridge change. The penalty window slices the token history handed to the existing Rust penalty functions ahead of the fused dispatch; the incremental SamplerState continues to serve only the full-history form (byte-identical, proven by the covering-window arm below), and a positive window takes the bounded rebuild path over at most repeat_last_n tokens. ignore_eos reuses the shared token-bias map (a -inf bias on the merged end-of-generation set at enqueue time), exactly upstream's EOG-bias mechanism.

What changed

• src/lib/mlxcel-core/src/generate.rs + sampling.rs: SamplingConfig::penalty_last_n, DRY_FULL_HISTORY, penalty_window, windowed penalty application, DRY sentinel arms and >= tier, needs_token_history window gating.
• src/lib/mlxcel-xla/src/sampler.rs + src/server/batch/xla_worker_admission.rs: the OpenXLA parity sampler mirrors the same window and DRY semantics (note for reviewers: this PR adds fields consumed by xla-iree-gated code; the OpenXLA feature compile job is the gate that sees those arms, though the xla crate itself compiles and passes its suite locally).
• Server flags on BOTH binaries: --top-nsigma/--top-n-sigma (default -1.0, sanitized), --xtc-probability, --xtc-threshold, --ignore-eos, repeatable -r/--reverse-prompt (merged into every request's stop set), --samplers/--sampler-seq/--sampling-seq (fixed-order validation failing startup with the accepted form in the diagnostic), plus --repeat-last-n semantics and --dry-penalty-last-n sentinel mapping; /props reports the new defaults with the key-set contract updated.
• Native /completion: top_n_sigma, xtc_probability/xtc_threshold (schema limits enforced with a 400), ignore_eos, repeat_last_n wiring, and samplers (default order accepted in both upstream shapes, others refused with a 400); generation_settings echoes the new resolved values.
• Wire protocol: SerializableSamplingState::penalty_last_n with a -1 serde default so an older peer's frame keeps its full-history penalties; round-trip and legacy-frame tests. Note: an old frame carrying dry_penalty_last_n: 0 (the old full-history spelling) is read as disabled by a new decode node during a mixed-version handoff window; recorded here rather than silently ignored.
• Manifest: compat/llama-server/b10621/sampling-and-grammar.json (27 flips, 31 repoints, two now-fixed --dry-base divergence records cleared, field:seed kept deferred under feat(inference): complete the remaining b10621 sampling and grammar surfaces #1485 because its below-minus-one rejection divergence is real) and compat/llama-server/b10621/pin.json (adds feat(inference): complete the remaining b10621 sampling and grammar surfaces #1485 to the shard owner set, the mechanism the validator itself prescribes for splitting a remainder).
• Docs: docs/python-client.md, docs/llama-server-compat.md owner table.

Numerical validation (Qwen3-4B-4bit, greedy decoding, every position decided)

Arm	Result
Control: --repeat-last-n -1 (full-history sentinel) vs --repeat-last-n 100000 (window covering the whole run), repetition 1.3, 200 tokens	byte-identical
Control: --repeat-last-n 64 with all penalties at identity vs no flags	byte-identical
Control: --dry-multiplier 0.8 --dry-penalty-last-n 0 vs DRY off	byte-identical (0 disables)
Treatment: --repeat-last-n 64 vs full history, repetition 1.3, 200 tokens	diverges first at word 76 of 164 (53 percent of subsequent words), exactly where the history first exceeds the window
DRY full-history sentinel vs DRY off	diverges (DRY active), arithmetic pinned by fixed-logit unit tests including the new base^0 tier

The teacher-forced forward path is untouched: no model or kernel file changes, so the logit-trace forward comparison is identically zero by construction; every change is post-forward sampling arithmetic, validated by the fixed-logit unit tests and the greedy real-checkpoint arms above.

Review follow-up (second commit)

A review pass against the pinned b10621 source corrected three claims the first commit got wrong about upstream, plus hardening: request stop strings now REPLACE the server-wide --reverse-prompt list when non-empty (upstream's antiprompt fallback semantics) instead of merging; the native XTC fields now CLAMP into b10621's soft 0..=1 limits instead of answering 400; both server binaries reject a negative --dry-penalty-last-n at parse time exactly as b10621's CLI does (the offline generate CLI keeps -1 as a documented mlxcel-only spelling); --dry-base is sanitized at startup and the core plus xla DRY implementations mirror upstream's below-1.0 early-out and exponent cap; the OpenXLA admission path rejects ignore_eos with a diagnostic instead of silently stopping at EOS and the disaggregated gap is recorded beside the existing XTC note; HTTP dry_penalty_last_n clamps to INT32_MAX so a wire value cannot collide with the internal full-history sentinel; allow_hyphen_values became allow_negative_numbers per the documented parser policy; check_sampler_order runs before the model load (fails in 21 ms); and two hijacked doc blocks were restored to their functions. The corresponding manifest notes were rewritten to describe upstream accurately.

Test plan

• cargo test -p mlxcel-core --profile test-fast --features metal,accelerate --lib sampling::tests (98 passed, including the seven new window/DRY tests)
• cargo test -p mlxcel-xla --profile test-fast --lib (280 passed; the breaker test now isolates the breaker from the new >= tier)
• Root filters (penalty, dry, sampler_order, sanitize, request_options, native_completion, serving_protocol, props, validate): all green
• cargo test --profile test-fast --features metal,accelerate --test llama_compat_manifest (4 passed against both rebuilt binaries) and the llama_compat lib suite
• python3 scripts/ci/check_llama_compat_manifest.py (zero feat(inference): complete b10621 sampling and grammar semantics #1436 deferred entries remain, so closing this issue cannot redden the validator), check_cross_repo_refs.py
• cargo clippy (-D warnings) on mlxcel-core, mlxcel-xla, and the root lib/tests/bins; cargo fmt --all -- --check
• Real-checkpoint arms as tabled above

Closes #1436