feat(adapter): Ollama (local) preset via /v1/chat/completions OpenAI-compat

[romer8]

Summary

Adds a new provider preset Ollama (local) that targets Ollama's OpenAI-compat endpoint (/v1/chat/completions) through the Django proxy. Fixes the multi-round tool-call regression observed against qwen3:30b-a3b-instruct-2507-q4_K_M on local Ollama >0.16.2:

Error: 400 json: cannot unmarshal object into Go struct field
.messages.tool_calls.function.arguments of type string

This is the architectural fix #23 explicitly named in its follow-up:

"the right fix is switch the Ollama adapter to use /v1/chat/completions (OpenAI-compat endpoint) instead of the native /api/chat, since the OpenAI-compat path has a stable, OpenAI-spec wire contract."

Implemented as a new preset rather than swapping the existing ollama provider so Ollama Cloud (which is happy on /api/chat) is untouched. Local Ollama participants pick "Ollama (local)" from the dropdown.

Why

Ollama's native /api/chat flipped the inbound shape of tool_calls[].function.arguments between versions — object on <=0.16.2, string on >0.16.2. PR #22 stringified universally → broke 0.16.2's chat-template parser → PR #23 reverted. The two versions can't be satisfied by a single static wire format. /v1/chat/completions follows OpenAI spec stably across all Ollama versions (arguments always stringified).

Changes

• storage/llmProviderStorage.js — new ollama_local preset.
• engine/adapters/ollamaLocal.js (new) — OpenAI SDK wrapper with baseURL = ${location.origin}/apps/tethysdash/ollama-proxy/v1 and defaultHeaders carrying x-csrftoken, x-ollama-host, x-ollama-key so the Django proxy can route to the user-configured Ollama host.
• engine/index.js — ollama_local → ollamaLocalStreamChat in PROVIDER_ADAPTERS.
• helpers/index.js — listModels ollama_local branch reuses the existing /api/tags + /api/show discovery flow (same x-ollama-host routing).
• components/LLMProviderPanel.jsx — show URL field for ollama_local alongside ollama/custom.
• engine/adapters/ollamaLocal.test.js (new) — 9 tests including the regression catcher PR Revert "fix(engine): stringify tool_call arguments for Ollama /api/chat wire format (#22)" #23 named ("the test that would have caught this regression"): multi-round messages → second-round request's prior assistant tool_calls[0].function.arguments must be a string.

Requires

Companion controller in tethysapp-tethys_dash adding @controller(url="tethysdash/ollama-proxy/v1/chat/completions/") to proxy /v1/chat/completions upstream. Without it, this preset returns 404.

Test plan

• npx vitest run engine/adapters/ollamaLocal.test.js → 9/9 passed
• npx vitest run → 669/669 passed (no regressions)
• npm run build → dist rebuilt clean
• After the tethysdash companion lands, manual smoke against local Ollama with the original failing prompt: List available dates for cfe_nom model should now succeed and the assistant reaches the tool dispatch
• Workshop image rebuild + republish → participants see the new preset in the dropdown