feat(inference): complete b10621 sampling and grammar semantics #1436

Parent

• epic(server): close llama-server b10621 compatibility gaps #1431
• Related: feat(sampling): add locally typical sampling (typical_p) #1377

Problem

Beyond the low-risk aliases in #1430, b10621 exposes sampling algorithms and controls that mlxcel does not implement or maps differently: typical_p, top-n-sigma, dynamic temperature, Mirostat, adaptive-p, XTC edge semantics, sampler ordering, ignore-EOS, grammar/lazy-grammar triggers, preserved tokens, multiple completions, and native cache/progress fields.

Scope

• Use feat(sampling): add locally typical sampling (typical_p) #1377 for typical_p and add the remaining algorithms as independently testable sampler stages.
• Match b10621 value domains, disable sentinels, defaults, and ordering semantics, including XTC probability/threshold behavior.
• Wire repeat_last_n / repetition_context_size into the shared repetition, frequency, and presence penalty history; the field is currently parsed but not applied to the core sampler, and fix(server): align llama-server b10621 aliases and defaults #1430 intentionally stops before this arithmetic change.
• Correct dry_penalty_last_n sentinels: b10621 uses 0 to disable DRY history penalties, while mlxcel currently resolves 0 as full history; preserve an explicit full-history form without conflating it with upstream disablement.
• Add explicit sampler-sequence configuration and validate incompatible/repeated stages.
• Implement ignore-EOS, n completions, time/token limits, keep/discard/cache controls, and response-field selection where supported.
• Implement direct GBNF and lazy grammar triggers/preserved-token behavior or reject unsupported grammar forms before generation.
• Keep OpenAI-shaped and native request aliases synchronized without silently dropping recognized fields.

Acceptance criteria

• A field-by-field schema test covers all b10621 sampling request fields, aliases, defaults, and disable sentinels: the pinned manifest plus tests/llama_compat_manifest.rs and src/server/llama_compat_tests.rs hold every entry's state, spelling, default, and declaration in both directions; every entry this issue owned left deferred (26 supported, 3 not_applicable with tested rejection, the remainder split to feat(inference): complete the remaining b10621 sampling and grammar surfaces #1485 with the shard owner set extended).
• Deterministic sampler unit tests use fixed logits and seeds for each delivered stage (penalty window, DRY sentinels and off-by-one tier, top-n-sigma/XTC/ignore-eos surfaces) and for configured ordering (--samplers / --sampler-seq / native samplers accept exactly the fixed b10621 default order and reject anything else). The unimplemented algorithm stages (Mirostat, dynatemp, adaptive-p) moved to feat(inference): complete the remaining b10621 sampling and grammar surfaces #1485.
• Repetition, frequency, and presence penalties use exactly the configured recent-token window, with 0 disabling the shared penalty stage and the full-history sentinel (-1) preserving existing behavior.
• DRY lookback 0 disables the DRY history stage, 64 is the server default, and the documented full-history sentinel (DRY_FULL_HISTORY, spelled -1 on the CLI flags) remains available without changing the meaning of upstream-shaped requests. The >= off-by-one tier and the dry_base < 1.0 fallback also landed.
• Grammar tests cover eager/lazy operation, triggers, preserved tokens, malformed input, and tokenizer edge cases. Moved to feat(inference): complete the remaining b10621 sampling and grammar surfaces #1485 together with the grammar flags and fields; mlxcel's constrained decoding currently covers only the OpenAI-shaped response_format.
• Unknown or unsupported semantic fields return actionable errors rather than being ignored for every surface this issue claimed (sampler order, XTC domains, native rejections); the fields split to feat(inference): complete the remaining b10621 sampling and grammar surfaces #1485 remain deferred there rather than claimed.
• Existing OpenAI request behavior remains compatible where field names overlap: repetition_context_size (alias repeat_last_n) kept its spelling and now actually reaches the sampler, and every default-off surface preserves the byte-identical baseline.

Numerical validation

For any implementation that changes arithmetic or kernel selection, run the teacher-forced logit-trace procedure from docs/benchmarks.md at realistic decode and verify widths and report disagreement at decided positions. Byte identity or perplexity alone is insufficient.

Split: the remaining algorithms and value-domain ports (Mirostat, dynatemp, adaptive-p, grammar surfaces, logit-bias, dry-sequence-breaker strings, min_keep, n_probs, post_sampling_probs, backend_sampling, and the temp/top-k/top-p/seed default-resolution divergences) moved to #1485, which now co-owns the sampling-and-grammar shard.