[VickyXAI]

What changed

This ports Franklin's src/free-models.ts and src/free-models.constants.ts and removes the free model ids this fork still had hard-coded across about 15 call sites. Those ids left the gateway catalog: nemotron-nano-9b-v2, nemotron-nano-12b-v2-vl and mistral-nemotron in the 2026-08-30 pool rotation, and qwen3-coder-480b and llama-4-maverick earlier. The gateway answered for them with a substitute, so these calls never got the model they named.

• New module: FREE_DEFAULT_MODEL, FREE_PREFERENCE_ORDER, freeChain, isFreeModelId, freeVisionModel and FREE_POOL_CONTEXT_FLOOR, matching Franklin. isFreeModelId reads the live gateway cache plus the static lists, because this fork has no @blockrun/model-catalog dependency.
• Literals replaced with the shared default or chain: subagent, moa, evaluator, verification, turn-analyzer, compact, llm static fallback, media-router, brain/extract, learnings/extractor, telegram, proxy fallback, plugin-sdk workflow tiers, and the picker. The picker re-exports FREE_DEFAULT_MODEL for existing importers.
• Router:
  • The free profile, getFallbackChain('free') and pickFreeFallback now use the live free chain, filtered by the runtime kill-switch. The chain no longer differs by category.
  • The quarantine list comes from free-models.
  • The tier classifier now uses FREE_DEFAULT_MODEL with a token budget a reasoning model can finish in, and parses the last tier word on the last line.
• Vision:
  • No free model is treated as vision-capable by default. nano-12b-v2-vl is gone, and the gateway refuses images for nano-omni on /v1/messages.
  • A free model is never swapped to a paid vision model, in either the agent loop or the proxy. A free-profile vision turn tells the user it is answering from text only.
  • A free proxy request never falls back to a paid model.
• Free checks: isFreeModelId replaces the startsWith('nvidia/') check in the subagent spend gate, the /model paid warning, compaction and isWeakModel.
• Timeouts and token budgets: the turn-analyzer moves to 1024 tokens / 8s, matching Franklin. The media-router moves to 2048 tokens / 10s. That one is my own sizing, not Franklin's, because Franklin removed media-router. It replaces maverick's 384 / 3.5s, which would cut a reasoning model off mid-answer.
• tokens.ts: live free ids are sized to FREE_POOL_CONTEXT_FLOOR, as in Franklin.
• Dead DeepSeek ids: nvidia/deepseek-v4-flash and nvidia/deepseek-v4-pro are in no free set or routing order, and a test pins this.
• Tests: free routing pins are updated, and a new guard fails if any runtime default in src/ names a retired free id.

Checks

npm test: 393/393 pass. npx tsc --noEmit clean, node dist/index.js --help OK, brand-numbers check up to date.

No version bump, no release.