fix(nvidia): add availability sync and remove 58 stale models

[fastreds]

Verify nvidia provider models against the live NVIDIA NIM catalog endpoint (https://integrate.api.nvidia.com/v1/models). This PR:

• Adds packages/core/src/sync/providers/nvidia.ts, an availability sync module that preserves TOMLs for models still served, removes TOMLs no longer in the catalog, and files issues for missing catalog models. Registered in index.ts and the direct sync group.
• Removes 58 stale nvidia-provider TOMLs that are no longer listed by the live API (e.g.
  vidia/synthetic-video-detector ->
  vidia/ai-synthetic-video-detector, deepseek-ai/deepseek-v4-flash -> deepseek-v4-flash-0731, qwen/qwen3-coder-480b-a35b-instruct, moonshotai/kimi-k2-instruct-0905, microsoft/phi-4-*, mistralai/mixtral-8x7b-instruct, etc.).

Going forward the hourly sync keeps the provider fresh automatically. �un models:sync nvidia --dry-run reports 58 removed / 0 created.

Source of truth: https://integrate.api.nvidia.com/v1/models

[github-actions]

Action items

• [high] [violation] packages/core/src/sync/providers/nvidia.ts:44 - Check: Availability syncs must preserve hand-authored TOMLs and base_model factoring (see sync.md / OpenAI availability pattern). Why: translateModel returns context.existing(model.id) (base-resolved merge), not context.authored. The runner then re-validates and formatTomls that expanded shape. For every still-served model with base_model (e.g. openai/gpt-oss-120b, nvidia/nemotron-3-nano-30b-a3b, minimaxai/minimax-m3), the next hourly bun models:sync nvidia will rewrite override-only files into full inlined copies and churn diffs. The comment claims “preserve the authored TOML verbatim,” but this path does not. Action: Mirror openai.ts / ofox.ts: look up context.authored(model.id), return that shape (or an OpenAI-style preserveAuthoredModel helper), and only use resolved existing if you truly need merged fields.
• [medium] [violation] packages/core/src/sync/providers/nvidia.ts:32 - Check: skipCreates missing-model tracking must match documented behavior (sync.md Missing-model GitHub issues). Why: Module header, translateModel comment, and the PR body say missing catalog IDs open issues / are deferred as missing-model issues, but trackMissingModels: false disables issue opens, and there is no skippedNotice. New NVIDIA IDs are dropped with no issue, notice, or create—silent catalog drift. Action: Either set trackMissingModels: true (and keep skipCreates: true) if the public NIM list should drive missing-model issues, or keep false and fix the comments/PR text; in the latter case add a skippedNotice so remote-only IDs are still reported in the sync report.