Match runtime decoding precedence, not only API fields

When recreating a custom generative inference service, trace where decoding parameters are actually consumed. An API may accept per-request sampling fields while the engine resolves one process-wide policy from checkpoint metadata and silently ignores those fields. Substituting greedy decoding can suppress natural stop-token behavior and turn normal outputs into hard-cap completions. Preserve the original precedence rules, log resolved values at startup, and verify with both cap math and repeated outputs. This does not imply sampling always improves stop behavior; it applies only when the prior runtime truly used checkpoint defaults.

Applies when: Migrating a custom model server while preserving its external API contract. Keywords: generation config sampling greedy EOS hard cap runtime parity