feat(ai): add qwen3.8-max to the prime inference catalog

[eliebak]

Summary

qwen/qwen3.8-max is live on Prime Inference but doesn't show in current prime-agent installs. Note the picker list is regenerated at build time (packages/ai build runs generate-models, and build-binaries.yml/ci.yml run npm run build), so the next release would pick the model up in the long tail automatically. What this PR adds on top of that:

• Featured pin: adds qwen/qwen3.8-max to PRIME_INFERENCE_FEATURED_MODELS so it's pinned above the long tail in /model (alongside qwen/qwen3-max) — this is a source change that no rebuild picks up automatically.
• Fresh committed baseline: regenerates the snapshot's prime-inference section. The committed snapshot is the fallback the generator keeps when the OpenRouter metadata fetch fails at build time, and it's what dev builds start from — a stale baseline would silently drop these models in that path. The regen is scoped to the prime-inference section only; all other provider sections are byte-identical to main to avoid unrelated build-environment churn.

Prime Inference section changes

• Added (new on the live catalog since the last committed regen, chore(ai): regenerate model catalog snapshot #561 on 2026-07-31): qwen/qwen3.8-max (featured, 1M context, 131k max tokens, vision + reasoning), deepseek/deepseek-v4-flash-0731, poolside/laguna-s-2.1, poolside/laguna-xs-2.1
• Removed (no longer served): mistralai/mistral-small-24b-instruct-2501, mistralai/mistral-small-3.2-24b-instruct, nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B
• Refreshed max-token limits for a handful of existing routes from current catalog/OpenRouter metadata

All four added models are on the unauthenticated public catalog (verified against https://api.pinference.ai/api/v1/models), so CI's credential-less build sees them too. The QWEN3.8 MAX display name follows the existing convention for this provider (QWEN3 MAX, QWEN3.5 397B A17B).

Testing

• packages/ai: test/prime-inference-models.test.ts passes (10/10), tsgo --noEmit clean
• packages/coding-agent: test/model-resolver.test.ts passes (39/39)
• Verified the diff touches no provider section other than prime-inference

🤖 Generated with Claude Code

---

Note

Low Risk
Catalog metadata and featured-model list updates only; no auth, request, or runtime logic changes.

Overview
Pins qwen/qwen3.8-max in PRIME_INFERENCE_FEATURED_MODELS so it appears above the long tail in the model picker.

Also regenerates the committed prime-inference snapshot: adds qwen/qwen3.8-max, deepseek/deepseek-v4-flash-0731, and the Poolside Laguna 2.1 models; drops three retired routes; and refreshes a few maxTokens/cost values from the live catalog.

^{Reviewed by Cursor Bugbot for commit 37c6d18. Bugbot is set up for automated code reviews on this repo. Configure here.}