fix(inference): align context, batching, RoPE, and runtime override semantics #1450

Parent

• epic(server): close llama-server b10621 compatibility gaps #1431

Problem

Several common b10621 options affect model arithmetic or context management independently of the GGML hardware backend: context size, logical/physical batch size, kept prompt tokens, SWA cache sizing, RoPE scaling/base/frequency, YaRN controls, escape processing, tensor checks, and model metadata overrides. mlxcel currently shares some spellings but does not establish equivalent defaults or effects.

Scope

• Audit and align --ctx-size, --batch-size, --ubatch-size, --keep, and --swa-full value domains, defaults, and truncation behavior.
• Implement compatible RoPE scaling, scale, frequency base/scale, and YaRN parameters only for model families whose Rust implementation can honor them.
• Align escape/no-escape processing at CLI prompt boundaries without changing JSON request escaping.
• Classify performance-timing, tensor-check, metadata-override, and operation-offload flags as supported, not applicable, or deferred.
• Align warmup/no-warmup startup behavior and prove whether the empty warmup run allocates the same caches and surfaces the same startup failures.
• Reject model-family or checkpoint combinations that cannot safely apply a requested runtime override.

Acceptance criteria

• CLI and canonical environment precedence tests cover every common runtime option and sentinel. (RoPE/YaRN group: src/cli/rope_args_tests.rs, tests/llama_compat_manifest.rs. The remaining runtime options in this shard were audited and their divergences recorded in compat/llama-server/b10621/runtime-and-context.json rather than claimed.)
• Resolved context and batch settings are observable in startup diagnostics and native props. (/props now reports n_ctx, n_batch, n_ubatch, n_batch_decode, n_kv_max; startup logs the resolved per-slot context, slot count and prefill chunk.)
• Truncation/keep/SWA tests cover full context, sliding windows, rotating cache, and MoE/VLM model families where applicable. (Not implemented. --keep, --swa-full, --context-shift, n_keep and n_discard stay deferred with their observable divergences recorded; the retained prefix is still the fixed 4-token attention sink and sliding-window families build their own RotatingKVCache with no server-side reach into it.)
• RoPE/YaRN overrides are applied before cache allocation and are rejected for unsupported architectures. (Applied inside model construction, which precedes make_caches; an override that reaches no RoPE seam refuses to serve, verified by an applications counter rather than a hand-maintained architecture list.)
• No arithmetic-affecting option is accepted without a behavioral or numerical test. (Teacher-forced logit traces over 8192 positions, plus two per-process integration tests and 24 unit tests.)
• Warmup enablement, disablement, latency boundary, and startup-failure propagation are covered with a real checkpoint. (Out of this shard: --warmup is owned by model-source.json / fix(server): translate model-source flags and reject unsupported GGUF semantics #1434, so this chain cannot edit its entry.)

Numerical validation

Use the teacher-forced logit-trace procedure from docs/benchmarks.md at the configured prompt/decode widths and report disagreement at decided positions for every new scaling or batching path.