feat(ai): add Ollama Cloud provider

[thisisjun786]

Summary

• add Ollama Cloud as a built-in dynamic provider using OLLAMA_API_KEY
• discover tool-capable models through /api/tags and /api/show, with bounded inspection concurrency and last-known cache preservation
• expose thinking, vision, context-window, and OpenAI-compatible request metadata in the model catalog
• preserve explicit local Ollama models.json catalogs while keeping Radius and other provider-owned refresh paths intact
• document setup, model selection, environment configuration, and the local/Cloud compatibility rule

Why this belongs in core

Provider-owned catalog refresh is composed before extensions and models.json overlays. The runtime previously replaced a built-in provider's own refresh implementation with the generic remote catalog wrapper, so an extension could not make Ollama Cloud appear as a first-class provider or preserve its dynamic catalog lifecycle.

Verification

• npm run check
• npm run build:npm
• CI=1 npm test with provider credentials unset
  • scripts: 72 passed
  • agent: 303 passed, 1 skipped
  • AI: 1,526 passed, 804 skipped
  • coding-agent: 6,015 passed, 33 skipped
  • pty: 44 passed
  • senpi-codemode: 466 passed, 6 skipped
  • server: 3 passed
  • TUI: passed (exit 0)
• Senpi QA mock OpenAI-completions tool loop: 4/4 passed
• Senpi CLI smoke: 8/8 passed
• manual Ollama Cloud smoke: dynamic catalog discovery succeeded and a Kimi K3 text turn completed without an event-order error

Compatibility and failure handling

• an explicit ollama model catalog remains local and does not trigger Cloud discovery
• custom Radius gateways with explicit models still run their dynamic refresh
• one stale /api/show result does not discard healthy models
• failed or aborted refreshes cannot replace the last-known usable catalog
• upstream error bodies and status text are not included in surfaced errors

---

Summary by cubic

Adds a built-in Ollama Cloud provider with OpenAI-compatible /v1 support using OLLAMA_API_KEY, featuring dynamic, cached model discovery from /api/tags and /api/show. The runtime now preserves provider-owned refresh so both Ollama Cloud and Radius keep their dynamic catalogs.

• New Features

  • New ollama provider (Cloud) with /v1/chat/completions.
  • Discovers tool-capable models via /api/tags + /api/show; derives thinking, vision, and context window.
  • Bounded inspection concurrency; preserves last-known usable models on partial failures; aborts don’t overwrite cache.
  • Reasoning “max” clamps to Ollama’s highest supported “high” wire tier.
  • Zero-cost pricing metadata (no fabricated per-token rates).
  • Recognizes OLLAMA_API_KEY; docs and CLI help updated; default model set to qwen3.5:397b.
  • Runtime no longer wraps built-ins that implement refreshModels, keeping Ollama Cloud and Radius dynamic catalogs intact.

• Migration

  • Set OLLAMA_API_KEY, then run: senpi update --models and select an Ollama Cloud model (e.g., qwen3.5:397b).
  • Existing local ollama entries with explicit models in models.json are unchanged and skip Cloud discovery.

^{Written for commit fb26e7c. Summary will update on new commits.}