v0.12.245 — blockrun catch-up: deepseek-v4-flash EOL, 10 new models, 6 repriced

Catch-up sync to blockrun — the drift was three blockrun releases deep. Catalog is now 70 chat-visible / 5 free, matching blockrun exactly.

Removed

• free/deepseek-v4-flash — NVIDIA published EOL (HTTP 410) this morning; the whole nvidia/deepseek-* family is dead upstream (blockrun #367). Dropped from the picker (6→5 free), the FREE_MODELS cascade (8→7), and router-core's eco SIMPLE chain. Pins naming the model stay routable via the gateway redirect; generic shorthands follow blockrun's retarget to free/gpt-oss-120b. It was the last 1M-context free model.

Added — 10 models blockrun shipped that were never mirrored

GPT-5.6 Sol/Terra/Luna Pro tiers, Gemini 3.6 Flash, Gemini 3.5 Flash Lite, Qwen3.7 Plus/Flash, Tencent Hy3, Xiaomi MiMo-V2.5 Pro, Nano Banana 2 (image, $0.09), Seedance 2.0 Mini (video, $0.079/s, 720p + synced audio). toolCalling LIVE-VERIFIED on all seven flagged chat models via structured tool_calls probes through the live gateway.

Fixed — 6 stale prices

gpt-5.6-terra 2.00/12.00 + luna 0.20/1.20 (OpenAI's 07-30 cut), deepseek-chat/reasoner 0.14/0.28, glm-5 1.00/3.20, gemini-3.5-flash 1.50/9.00 (was under-logging by 3× and skewing maxCostPerRun accounting). Charges are always server-dictated via 402 — these feed telemetry and the cost-cap gate.

Full details in CHANGELOG.md.