v0.4.0

The open-models port. Every model seam now runs on any pydantic-ai provider — OpenRouter first — and the reference deployment moved to open-weight models behind this repo's own instruments: the filing golden passes every bar on deepseek-v4-flash at roughly $0.008 per capture, the QA golden scores 1.00 on every axis on glm-5.2, and the retrieval golden holds recall@5 at 0.969 with the same embedding model served through the new host. No gate, ACL, fence or eval moved: the seams did.

Added

• ask takes the two-form model convention — a bare ANSWER_MODEL stays the OpenAI Responses API; a provider-prefixed pydantic-ai id authenticates with that provider's own key — and openrouter joins the librarian's key preflight and pricing table (#107)
• the embedder speaks to any OpenAI-compatible /embeddings host (EMBED_BASE_URL, EMBED_API_KEY, EMBED_MODEL), with the recorded-model rule keeping every query in its index's vector space; the worker boot strips the embed credential like the OpenAI one (#108)
• two-form vision OCR: a provider-prefixed VISION_MODEL transcribes poppler-rasterized page images through pydantic-ai — bounded pages, a spoken cut — while the bare Gemini form stays byte-for-byte what it was (#109)

Fixed

• an empty model-chosen search query is a repairable refusal handed back to the asking agent, instead of a provider 400 that crashed the whole ask (#114)
• the audit sweep over the port (#114): the OpenAI key never travels to a non-default embedding host; a prefixed VISION_MODEL missing its provider's key refuses naming that key instead of advising a requeue that could never work; the rasterizer carries timeouts and a pixel bound against raster-bomb pages; the provider→key table lives once, in the kernel; the worker boot says out loud when one credential wears two names. Deferred halves filed as #110–#113
• account schemas decode nested structures a provider's tool-calling stringifies — the defect that filed meetings with decisions empty on routes returning nested lists as JSON strings (#114)