fix(free): replace the delisted nemotron-3-nano-30b, and document the live free lineup

[VickyXAI]

This repo allows squash merges only. Squash-merge this PR.

Depends on BlockRunAI/router-core#3. The router_core port is re-synced to that PR's head e38958b, and the decision fixture is its regenerated snapshot. Land router-core#3 first. If it lands under a different SHA, only the pin references in three docstrings change; the code and fixture do not.

Sources, fetched 2026-09-29: the $0 chat set in blockrun.ai/api/v1/models is nvidia/nemotron-3-nano-omni-30b-a3b-reasoning, nvidia/nemotron-3.5-lightning, nvidia/llama-3.2-11b-vision, nvidia/nemotron-3-ultra-550b, cohere/north-mini-code and poolside/laguna-xs-2.1. brand/numbers.json has models.free = 6. nvidia/nemotron-3-nano-30b was delisted on 2026-09-08, and the gateway redirects it to nano-omni. gpt-oss-120b/20b were retired on 2026-09-03 (available:false, redirected).

Code

Where	Old	New	Why
router_adapter.FREE_TIERS SIMPLE primary	nvidia/nemotron-3-nano-30b	nvidia/nemotron-3-nano-omni-30b-a3b-reasoning	nano-30b is no longer in the catalog. The adapter drops uncatalogued candidates, so SIMPLE was silently opening on its first fallback. nano-omni is the gateway's own redirect target for nano-30b and the same family and size (30B-A3B), so it is the model these slots were actually served by. The 2026-08-31 reason for excluding it (it answered as nano-30b) went away with nano-30b. On 2026-09-29 a model-echo probe got nano-omni's own NIM deployment back.
FREE_TIERS MEDIUM / COMPLEX / REASONING fallbacks	nano-30b	nano-omni	Same. Each tier keeps its depth of ≥3 distinct models.
router_core/config.py eco SIMPLE fallback[0]	nano-30b	nano-omni	Port of router-core e38958b
router_core/model_capabilities.py	nano-30b entry; nano-omni supports_vision: True	nano-30b removed; nano-omni supports_vision: False with the upstream override comment	Port. Without it, eco SIMPLE image turns would start landing on a model whose image path fails real probes.
tests/unit/router_core_decisions.snapshot.json	upstream 5ee7c23	upstream e38958b, verbatim	The only decision change is one candidate id. The Python port reproduces it exactly (test_router_core_snapshot.py passes).
router_core/__init__.py, config.py, test_router_core_snapshot.py	pin 5ee7c23	e38958b
tests/unit/test_router_adapter.py	FREE_MODELS lists nano-30b	nano-omni
tests/unit/test_router_adapter.py	—	test_names_only_models_in_the_live_free_catalog	The existing check compared FREE_TIERS with the hand-kept FREE_MODELS, which can't catch a delisting. The new test pins the table to the published $0 set and names nano-30b. I checked that it fails on the old table.
tests/unit/test_routing_parity.py catalog fixture	nano-30b row	nano-omni row
client.py chat_completion_stream docstring	nvidia/deepseek-v4-flash, fallback nvidia/llama-4-maverick	nvidia/nemotron-3.5-lightning, fallback nano-omni	Both old ids are retired

README

Section	Old	New
Available free models	step-3.7-flash, mistral-nemotron, nano-omni, nemotron-nano-9b/12b, gpt-oss-120b/20b ("all NVIDIA-hosted")	The six live ids exactly as /v1/models lists them, a note that neither vision-catalogued free model handles images reliably, and the retired list extended with the 08-30, 09-03 and 09-08 retirements. The gpt-oss privacy note is removed (those models are retired).
free routing-profile row	"Step 3.7 Flash, Mistral Nemotron, Nemotron Nano Omni / 9B / 12B VL"	The five models FREE_TIERS routes across, plus Ultra 550B as direct-call only (it is deliberately not in the table)
Hero line	"8 fully-free NVIDIA-hosted models — DeepSeek V4 Flash … Llama 4 Maverick …"	br:models.free marker plus the six names
Try-it-free and funding examples	nvidia/step-3.7-flash	nvidia/nemotron-3.5-lightning; result.model example is nano-omni
NVIDIA section	08-12 free rows	The four live nvidia/* free models, with the non-NVIDIA pair cross-referenced
FAQ	"11 NVIDIA-hosted models"	br:models.free marker

Left alone on purpose

• nemotron-3-ultra-550b stays out of FREE_TIERS. It was excluded on 2026-08-31 because it answered as another model, and today's probe got a 429, so there is no fresh evidence either way.
• examples/sweep_all_chat_models.py hard-codes retired ids. It is a historical sweep script, not a list the SDK uses.
• Testnet openai/gpt-oss-* rows are paid testnet models, not the free tier.
• CHANGELOG is written at release time.

Checks (throwaway venv, Python 3.13, .[dev,solana,anthropic] as in CI)

• pytest tests/unit: 984 passed
• black --check .: pass
• ruff check .: pass
• node scripts/sync-brand-numbers.mjs --check: up to date
• mypy blockrun_llm: 256 errors, identical to main (mypy is not a CI step)

🤖 Generated with Claude Code