Fallback chains depend on models hidden from /v1/models, and the fix path now spans three repos

[VickyXAI]

Follow-up from the Router Core extraction review. Not a bug today — everything below resolves correctly right now. It is a note about how much slower recovery got.

The dependency

Several load-bearing routing targets are hidden: true in blockrun's catalog — deliberately excluded from /v1/models, but still available: true and therefore callable by direct ID:

Target	Where it is load-bearing
nvidia/gpt-oss-120b	ultimate free backstop, both AGENTIC tiers (config.ts, set by 6a790eb)
nvidia/gpt-oss-20b	head of the free coder fallback chain
nvidia/qwen3-next-80b-a3b-instruct	Franklin free-profile default
moonshot/kimi-k2.5 / k2.6 / k2.7	Franklin VISION_MODELS (re-added in Franklin#122)

They work only because available: true. The catalog hides them for a real reason — NVIDIA's free tier reserves the right to use prompts for service improvement — so the flag they depend on is one that gets revisited.

This exact class already caused breakage before: hidden: true strips a model from /v1/models and silently breaks router primaries targeting it.

What the extraction changed

Before, retargeting a dead free model was one commit in ClawRouter. Now the chains live here, pinned by immutable commit SHA in two consumers:

router-core commit → repin ClawRouter → release → repin Franklin → release

6a790eb ("retire unavailable seed fallback") is that path being walked for real, and seed-oss-36b had already been retargeted once before that. This is not a rare event.

Worth considering

• A host-supplied override for fallback chains, so a consumer can retarget a dead rung without waiting on a core release (capabilities and pricing are already injected this way).
• Or: treat a hidden: true model as ineligible for an ultimate backstop slot specifically, since that rung is the one with nothing behind it.
• At minimum, document that chain entries must be available: true regardless of catalog visibility, so the invariant is explicit rather than discovered.

Filed against core because that is where the chains now live; the consumer-side follow-up is the repin cadence, not a code change.