fix(core): stop offering deprecated models in fallback lists

[yihanzhu]

Summary

fallbackModels is a recommendation, not an inventory: everything in it lands in the catalog as available and default-capable, and its head is what a new connection defaults to and what the connection test probes. toolCallingModelIds filters on tool-calling capability only, so six providers still offered 28 models their vendors have deprecated.

This applies the call-site lifecycle filter #3324 established for ZenMux to xiaomi, nvidia, mistral, togetherai and deepinfra, and drops the one deprecated id from openai's hand-written list. That empties PROVIDERS_WITH_DEPRECATED_FALLBACKS, so the recorded boundary goes away with it and the remaining assertion becomes the unconditional invariant — which is what its comment says it was built to do.

Removal is from the offer only. An id a user already chose still sends (authorizeConnectionModel never reads this list), and all six do kind: 'protocol' discovery, so the endpoint's own catalog returns on the first /models fetch. Every list head is unchanged, so no connection default or connection-test probe model moves.

Reasoning for the offer-policy question the issue poses, and the measurements behind the notes below, are in #3355 (comment).

Fixes #3355

Review focus

• Two corrections to the issue body. The 28 is right as a registry count, but only 27 are user-visible: CURATED_CATALOG_FALLBACK_MODELS.openai shadows the registry list at model-catalog.ts:239, so gpt-4-turbo is in no picker today — removing it is hygiene for the non-catalog consumers (connection test, onboarding verify, CLI projection) and for the contract test, which reads the registry directly. And "discovery keeps only the fallback set" is the kind: 'fallback' branch; none of these six is on it, which is why the exclusion costs no reachability.
• Two product judgments worth naming. mistral loses its whole devstral line from the pre-credential list (all eight deprecated ids are devstral plus open-mistral-nemo), and xiaomi is the largest proportional cut, 6 → 3. Both come straight back from live discovery.
• The clause now sits at 13 call sites, so a new provider is wrong by default. Folding it into toolCallingModelIds would be a no-op today — of 46 snapshots, the only ones carrying a tool-capable deprecated id are the nine now filtered at their call site plus openai, google and the three xiaomi-token-plan entries, which all use hand-written lists that never reach the helper. Happy to roll that in here or leave it as a follow-up; the wrinkle is placement, since ollama-cloud's pre-filter is what makes a deprecated recommended id throw at module load rather than vanish.
• Known limit, pre-existing: the invariant sweeps CATALOG_PROVIDER_TYPES, which is 55 of 59 registry entries. xai-oauth, github-copilot, claude-subscription and openai-codex sit outside it; I checked and none carries a deprecated id today.

Verification

• Red state: with the registry change reverted, keeps deprecated snapshot models out of fallback lists fails naming all six providers and all 28 ids (togetherai 11, mistral 8, xiaomi 3, deepinfra 3, nvidia 2, openai 1)
• @maka/core 588 pass; @maka/storage 895 pass; @maka/runtime 3066 pass with one environment failure (claude-subscription-usage cannot reach the OAuth token endpoint in my sandbox — fails identically on unmodified main)
• Root lint, format:check and typecheck all pass
• Post-fix list sizes, zero deprecated remaining in each, heads unchanged: togetherai 19, mistral 22, xiaomi 3, deepinfra 56, nvidia 59, openai 3
• No protocol files touched, so RUNTIME_HOST_COMPATIBILITY_EPOCH stays put. model-pricing.generated.ts still carries rows for the removed ids and is deliberately untouched — it is generated and keyed independently.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Claude Code (Opus 5) — implementation, tests and the supporting measurements, under my direction and review. The commit carries a Generated-by trailer.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes — a connection created after this change sees a shorter pre-discovery model list; described under Summary above
• No

[jackwener]

Reviewed the exact head against the current base. The fallback lifecycle policy and scope are sound, the follow-up only corrects the contract comment, and I found no blocking correctness issue.