nemotron_asr: move options to the model spec

[LysanderdeJong]

Summary

• declare every existing nemotron_asr option in model_specs/nemotron_asr.json (schema_version: 1)
• validate session and request options with the spec-backed helpers; unknown request keys now fail
• generate the CLI help from the spec

First of three stacked PRs split from #681: this PR, then the diarizer safetensors output, then speaker-tagged ASR.

Why

The family validated session options with a hard-coded whitelist, read request options ad hoc, and hard-coded its CLI help. Adding an option needed three edits.

Changes

• model_specs/nemotron_asr.json: request options language, lookahead_tokens, max_tokens, keep_language_tags, return_timestamps, streaming; the seven existing session options.
• Sessions get the contract from the loader and call validate_spec_backed_session_options / validate_spec_backed_request_options in prepare(), run() and start_stream().
• inspect() returns the contract's CLI help. The loader stays custom, so metadata and the prompt-dictionary language list are unchanged.

Maintainer notes

• An undeclared request key, such as a typo, used to be ignored and now fails. Session keys were already checked by the old whitelist.
• The published Q8_0 GGUF embeds a legacy (pre-v1) spec. It keeps working through the existing fallback to the workspace v1 contract, which prints the usual legacy-spec warning. Repackaging it would embed the new options.
• Without any v1 contract (legacy embedded spec, non-deployment build, no model_specs/ reachable), nemotron_asr keeps main's behaviour: the old session whitelist, unvalidated request options and the hardcoded help. model_spec::find_model_contract() is the non-throwing lookup that makes this possible.

Validation

Windows 11, MSVC (VS 2026), CUDA 13.3.

• offline and streaming text and --words-out JSON are byte-identical before and after
• every existing option still works
• --request-option bogus=1 fails with unknown Nemotron ASR request option: bogus; --session-option nemotron_asr.bogus=1 fails with unknown Nemotron ASR session option: nemotron_asr.bogus
• tools/check_loader_catalog_sync.py passes

[0xShug0]

@LysanderdeJong The migration is generally in the right direction, but the PR breaks legacy GGUFs in non-deployment builds when no external v1 spec is available embeds a legacy model spec, but no current schema-v1 model contract was found. There are two approaches, and I'm fine with either. (1) Defer the migration for now and migrate all models together in a dedicated pass later, rather than breaking compatibility one model at a time. (2) Add a compataible layer: With a contract: use spec-backed validation. Without one: retain main’s existing whitelist. This way, downstream consumers still have a clean, single source of truth, while the compatibility mess stays inside the code.

[LysanderdeJong]

Thanks, went with option 2 in 8e59ca1:

• General (framework): model_spec::find_model_contract() / find_contract_spec_path() return nullopt instead of throwing when the GGUF embeds a legacy spec and no v1 contract exists. default_contract_spec_path() keeps its current behaviour and error.
• Model-specific (nemotron_asr): with a contract, spec-backed validation and help; without one, main's session whitelist, unvalidated request options and hardcoded help, unchanged.

Tested with the published Q8_0 GGUF from a working directory outside the repo (no model_specs/ reachable, non-deployment build): offline and streaming transcribe, unknown session keys still fail, help shows the legacy list. Inside the repo, v1 validation still rejects bogus=1. #686 and #681 are rebased on top.