v2.6.1

[2.6.1] — 2026-08-28 — Airframe 0.4.1 and Unified Chat Templating

Highlights

• Single chat-templating renderer. All paths (HTTP /api/generate, OpenAI
  /v1/chat/completions + /v1/completions, Anthropic /v1/messages, and the
  CLI generate) now render prompts through one prompt_render module that uses
  each model's real GGUF chat_template (Jinja via shimmyjinja), with a
  family fallback (ChatML/Llama3/Gemma) and a --raw escape for base/completion
  models. Previously instruct models were fed raw prompts or a coarse
  ChatML/Llama3/OpenChat heuristic — the root cause of garbled output on
  Qwen3/Gemma/etc.
• Real GGUF template sourced at load. ModelSpec.chat_template is populated
  from the GGUF header (airframe metadata), so Jinja-first rendering is automatic.
• shimmyjinja extended for real templates: slice-with-step (messages[::-1],
  used by Qwen3 thinking) + Gemma family fallback.
• Offline routing gate. tests/gguf_routing.rs validates template routing for
  every model in the models dir (no GPU, ~14s) — the cheap per-model "screen test".
• Legacy server retired. shimmy_server_gpu (a dead standalone GPU server with
  its own template renderer) removed — 1,847 lines deleted. See CHAT_TEMPLATING.md.

Docs

• New docs/CHAT_TEMPLATING.md — the canonical architecture reference.

Release

• Updated the Airframe dependency to 0.4.1.
• Published Linux x86_64, Windows x86_64, and macOS arm64 binaries.
• Corrected release artifact permissions and names so platform downloads are
  attached to the GitHub release.