Add a sync module for neuralwatt (public, metadata-rich /v1/models)

Context

The neuralwatt provider entry is hand-authored, and it drifted: #7482 had to add 8 missing models (glm-5.3-flash family, deepseek-v4.1-flash family, deepseek-v4-flash-speed, two qwen flex tiers), deprecate 8 retired ones (GLM-5.2 family, deepseek-v4-pro), and fix two stale output caps — all detectable from Neuralwatt's own catalog endpoint.

The endpoint is unusually sync-friendly

GET https://api.neuralwatt.com/v1/models is public (no auth) and returns per-model metadata covering nearly everything the catalog tracks:

• metadata.pricing: input_per_million, output_per_million, cached_input_per_million (USD/MTok), service_tier (standard/flex), flex_discount_multiplier (0.65)
• metadata.capabilities: tools, json_mode (→ structured_output), vision, reasoning, reasoning_effort, streaming, system_role
• metadata.reasoning: mandatory, default_enabled, supported_efforts, default_effort, accepted_efforts, effort_aliases — maps directly to reasoning_options effort values
• metadata.limits: max_context_length, max_output_tokens, max_images
• metadata.deprecated + deprecated_message (lifecycle signal)
• metadata.huggingface_id, display_name, description

Suggested shape

Follow the public-endpoint precedent of chutes.ts / ovhcloud.ts (no secrets needed):

• Pricing, limits, capabilities, and lifecycle (deprecated flag) are endpoint-authoritative.
• base_model resolution via huggingface_id/family against models/; unresolved IDs skip with a missing-model issue rather than creating inline entries.
• reasoning_options effort values from reasoning.supported_efforts. budget_tokens needs a policy call: portal docs say thinking_token_budget is accepted provider-wide, but deepseek-v4-flash rejects it with a 400 — so derive per-model or preserve authored values.
• Absence from the endpoint (or deprecated: true) → status = "deprecated" for previously-cataloged models.
• ID normalization: the API mixes separators (deepseek-v4_flash-speed vs deepseek-v4.1-flash); catalog convention is hyphens.
• -flex IDs are the same underlying model at the flex tier and should resolve to the base model's lab metadata.

Registering in packages/core/src/sync/index.ts would give neuralwatt the hourly automation PR like other synced providers, and this class of drift (new flash/flex variants appearing weeks before anyone hand-PRs them) would stop happening.

Happy to take this on if maintainers want it.

Start by reading the sync implementations in chutes.ts and ovhcloud.ts, then inspect packages/core/src/sync/index.ts and the existing entries under models/. Trace how endpoint metadata, base-model resolution, normalization, and deprecated models are represented. Done means neuralwatt is registered for recurring syncs and the endpoint's catalog changes produce correct model metadata and lifecycle updates.