fix(server): align llama-server b10621 aliases and defaults #1430

Background

The official llama.cpp nightly pointer resolved to b10621 on 2026-08-26. The macOS arm64 distribution reports build 10621 at commit c1d0e7a00; its archive SHA-256 is 429c8270608600188035e5e92f7d78dffb7900904fe7dd7e6a84f48068cd13cf. Comparing that binary's actual llama-server --help and route registration against mlxcel main found several low-risk compatibility defects that do not require new inference features.

Problems

• Canonical environment variables have drifted: upstream uses LLAMA_ARG_BATCH, LLAMA_ARG_UBATCH, LLAMA_ARG_SPEC_DRAFT_MODEL, LLAMA_ARG_THINK_BUDGET, LLAMA_ARG_LOG_FILE, LLAMA_ARG_CHAT_TEMPLATE, LLAMA_ARG_CHAT_TEMPLATE_FILE, LLAMA_ARG_ENDPOINT_METRICS, LLAMA_ARG_ENDPOINT_PROPS, and LLAMA_ARG_ENDPOINT_SLOTS; mlxcel either uses stale names or has no env binding.
• LLAMA_ARG_CACHE_REUSE is an integer minimum reuse chunk size upstream, but mlxcel currently parses it as a boolean alias for enabling the whole prompt cache. Values such as 0 therefore produce the opposite operational result.
• Upstream b10621 defaults to top_p=0.95, min_p=0.05, and dry_penalty_last_n=64; mlxcel currently defaults to 0.9, 0.1, and an all-history sentinel.
• The same request fields are spelled differently: llama-server accepts n_predict / max_completion_tokens, repeat_penalty, repeat_last_n, and reasoning_budget_tokens; mlxcel's OpenAI-shaped routes currently accept only their native spellings and silently ignore the llama-server names.
• llama-server accepts either one stop string or an array and accepts seed -1 as random; mlxcel currently requires an array and an unsigned seed.
• Safe route aliases are missing: GET /v1/health, POST /reranking, and POST /v1/reranking can use existing handlers without adding new inference behavior.
• --temperature is a llama-server alias for --temp but is not accepted by mlxcel-server.

Scope

• Add canonical b10621 spellings while retaining documented mlxcel legacy spellings as lower-precedence compatibility aliases where doing so is unambiguous.
• Stop interpreting LLAMA_ARG_CACHE_REUSE as a boolean prompt-cache enable switch. Accept 0 as the upstream no-extra-reuse default without disabling prompt caching, and fail startup clearly for unsupported positive values rather than silently changing a different setting.
• Align the three server sampling defaults and /props reporting with b10621.
• Add request deserialization aliases, one-or-many stop parsing, and negative-seed normalization without changing the inference algorithms.
• Map repeat_last_n to the existing repetition_context_size request field only. Applying that field as a finite penalty-history window changes arithmetic and is explicitly deferred to feat(inference): complete b10621 sampling and grammar semantics #1436 with the numerical validation required by docs/benchmarks.md.
• Add the three route aliases and include /v1/health in the unauthenticated health allowlist.
• Update CLI consistency tests, request/route tests, environment-variable documentation, and the changelog.

Out of scope

• New sampler algorithms, TLS/CORS transport features, router mode, LoRA hot-swap, native completion response expansion, slot persistence, or support for low-level GGML device flags.
• Changing --timeout, whose mlxcel meaning is currently a decode hang watchdog rather than llama-server's HTTP read/write timeout; that requires a separate migration issue.
• Pretending unsupported hidden flags work by accepting and ignoring values that have observable llama-server semantics.

Acceptance criteria

• Both mlxcel serve and mlxcel-server accept canonical environment variable names with CLI-over-env precedence and preserve explicitly documented legacy aliases at lower precedence.
• LLAMA_ARG_CACHE_REUSE=0 leaves prompt caching enabled, positive values fail with an actionable unsupported-setting error, and boolean values are no longer accepted as cache-enable aliases.
• Server defaults and /props report top_p=0.95, min_p=0.05, and dry_penalty_last_n=64.
• Chat and OpenAI completion requests deserialize the listed llama-server field aliases into their existing mlxcel request fields; a scalar stop string is equivalent to a one-element array, seed -1 resolves to nondeterministic sampling, and seeds below -1 are rejected.
• This issue does not make the parsed repetition window affect sampling; that arithmetic change remains tracked by feat(inference): complete b10621 sampling and grammar semantics #1436 and is not claimed as runtime parity here.
• GET /v1/health, POST /reranking, and POST /v1/reranking reach the existing handlers, with /v1/health public under API-key authentication.
• --temperature and --temp resolve identically on both server entry points.
• Focused CLI, request-schema, route, auth, and config-default tests pass; cargo fmt --check passes.

Reference

• Upstream release: https://github.com/ggml-org/llama.cpp/releases/tag/b10621
• Upstream argument definitions: https://github.com/ggml-org/llama.cpp/blob/c1d0e7a004015f23bc0233470b747b596f29b264/common/arg.cpp
• Upstream route registration: https://github.com/ggml-org/llama.cpp/blob/c1d0e7a004015f23bc0233470b747b596f29b264/tools/server/server.cpp