The change is well-targeted and internally consistent. The logic switch at openrouter.ts:181 from model.top_provider.context_length ?? model.context_length to model.context_length aligns OpenRouter with the other sync providers (cloudflare-workers-ai, ovhcloud, baseten, llmgateway all use model-level context), and the new regression test correctly captures the behavioral delta — it would fail under the old top_provider-first logic and passes under the new logic. Existing OpenRouter tests use equal context_length values or hit the isUnavailable skip path, so they remain valid. The xiaomi/mimo-v2.5.toml edit removes the stale 32K override so the provider entry inherits context = 1_048_576 from models/xiaomi/mimo-v2.5.toml, which matches what the new sync would emit. The top_provider.context_length schema field is now unused but retaining it is harmless.

• [low] packages/core/src/sync/providers/openrouter.ts:181 - The PR reinterpretation rests on the OpenRouter API field semantics (context_length = model context window vs. top_provider.context_length = per-provider serving limit), but neither the PR body nor a file comment cites OpenRouter's API/model docs to confirm that distinction. The xiaomi/mimo-v2.5 data change is corroborated by the matching models/xiaomi/mimo-v2.5.toml metadata, so this is not blocking; a first-party OpenRouter API reference stating what each context_length field represents would strengthen the broader behavioral change.