feat(daemon): Ollama backend wiring — settings + selector + UI (v0.1.1)

[OpenCircuitDev]

Summary

Track 1, item 1 from docs/AGENT_OPERATIONS.md — Ollama daemon wiring. v0.1.1 ships on this. backend = "ollama" is now a first-class configured choice end-to-end: settings (TOML + Tauri commands) → selector → API server → settings UI.

The Ollama adapter itself (crates/ocm-inference/src/ollama.rs, NDJSON over :11434) already existed and was bench-verified; this PR makes it constructible from settings rather than dead code.

What changed

Rust (crates/)

• ocm-daemon::settings — Backend enum (auto/llamacpp/vllm/ollama, serde lowercase, default Auto) + 3 new fields: backend, ollama_base_url, ollama_model. All new fields #[serde(default)] so a v0.1.0 settings.toml without these keys still parses (legacy-load test included).
• ocm-inference::selector — BackendKind::Ollama variant; DEFAULT_OLLAMA_BASE_URL = "http://127.0.0.1:11434"; new make_backend_for_kind(kind, inference_url, ollama_url, ollama_model) settings-driven constructor. Existing make_backend(base_url) preserved for back-compat. Auto-detect (detect_backend_kind) never returns Ollama — Ollama is opt-in (bridges to an external daemon; not something OCM auto-supervises).
• ocm-daemon::bootstrap — build_app_state resolves Settings.backend (Auto → existing detect; explicit → that kind), falls back to DEFAULT_OLLAMA_BASE_URL / ollama::DEFAULT_MODEL when URL/model unset, and calls make_backend_for_kind. The Ollama backend now flows through to AppState.backend, which is the abstraction the OpenAI-compat API server consumes — "API passes the configured model through" is satisfied at the bootstrap layer (no openai.rs change needed).

Tauri commands (get_settings / save_settings) need no change: they round-trip the whole Settings struct, so the new fields flow through automatically.

Frontend (frontend/)

• settings.ts — Backend type + 3 new fields on the Settings interface (mirror of Rust).
• settings/+page.svelte — <select> for backend (humanized labels: auto/llama.cpp/vLLM/Ollama) + two <input>s for Ollama URL and Ollama model. Placeholders show the daemon's native defaults (127.0.0.1:11434, llama3).

TDD

Followed the test-first discipline adapted for "CI is my compiler" (no local Rust toolchain on this machine).

Commit	Purpose	CI Run	Verdict
42004cc test(ollama): RED	13 new tests referencing symbols that don't exist yet	27378990831	❌ (expected — compile-fail on missing variants/fields)
cb61171 feat(ollama): GREEN	Minimal impl to satisfy tests	27379125343	❌ (clippy::derivable_impls — a lint, not a logic failure)
6963c3f fix(settings)	#[derive(Default)] + #[default] instead of manual impl	27379270693	✅ all 3 platforms
d8b126c feat(frontend)	TS type + Svelte UI	27379313091 Frontend CI	✅ 26s

The RED-then-GREEN audit trail is intentional — it proves each new test actually tested something that didn't exist, then turned green from the minimal impl.

New test coverage

• selector.rs (+4 tests): as_str() lockstep with Ollama::name(); default URL constant; make_backend_for_kind produces the expected backend name() for all three kinds.
• settings.rs (+5 tests): Backend::default() == Auto; lowercase TOML serde; full Ollama round-trip; forward-compat legacy-load (settings.toml without backend key still parses to Auto); all-variants TOML round-trip.
• bootstrap.rs (+4 tests): the headline assertion — Settings.backend = Ollama ⇒ AppState.backend.name() == "Ollama"; Ollama with no URL/model still constructs (defaults fill in); explicit LlamaCpp / Vllm override platform detect.

Test plan

• cargo clippy --workspace --all-targets -- -D warnings — green on ubuntu/macOS/windows (run 27379270693)
• cargo test --workspace — green on ubuntu/macOS/windows (run 27379270693; counts per crate visible in log)
• Frontend npm run check — 258 files, 0 errors, 0 warnings (local, then CI run 27379313091)
• Frontend npm run build — clean static build in 3.80s (local)
• Manual smoke (operator) — launch cargo tauri dev, pick Ollama in the new selector, point at a running ollama serve, send a chat. Not run here (no local Rust toolchain; per AGENT_OPERATIONS verification doctrine, CI is the compiler).
• Bench cross-check (operator) — the amnesia-ab sandbox already CONFIRMED Ollama+Mem0 end-to-end on 2026-06-11; no rerun needed unless the operator wants one against the wired path.

Out of scope (deferred per AGENT_OPERATIONS Track 1)

• Process supervision (item 2) — supervisor.rs is still #[allow(dead_code)]. Will land in a separate PR.
• Settings UI polish (item 3) — current UI shows all rows unconditionally. Conditional "show Ollama fields only when backend = ollama" is a UX nicety for a follow-up.
• mem0-v3-locomo activation (Track 2) — separate workload-packaging task.

🤖 Generated with Claude Code