feat: local Ollama + Claude-headless LLM/embedding backends (perception-rust)

[lollinng]

Rebased onto feat/perception-rust (the Python tree is deleted on this branch; this re-implements local-backend support in Rust).

What

Lets patcha run without the patcha-cloud login by making the embedding and LLM backends pluggable. All opt-in via config; defaults (fastembed embeddings, patcha-cloud LLM) are unchanged.

Backends

Embeddings — EMBEDDING_PROVIDER

• fastembed (default) — unchanged, on-device.
• ollama — Embedder gains an enum backend; the Ollama variant calls /api/embed via a blocking ureq client. Rationale: embed_one/embed_many are sync, run inside block_in_place/spawn_blocking, and are dropped inside async in the CLI paths — a client that owns no tokio runtime avoids both the block-in-runtime and drop-panic failure modes reqwest::blocking would hit. Fails fast if Ollama is unreachable or the model isn't 768-dim (the store's fixed sqlite-vec width).

LLM — LLM_PROVIDER (routed inside PatchaApiClient::chat_completion; the 8 call sites are untouched)

• patcha (default) — unchanged cloud path.
• ollama — OpenAI-compatible /v1/chat/completions, no auth.
• claude — Claude Code headless: shells out to claude -p --system-prompt … --model <opus> --output-format text via tokio::process, using the local Claude login (no API key). --system-prompt overrides Claude Code's default agent prompt with patcha's.

Config (env-driven, all optional)

EMBEDDING_PROVIDER, LLM_PROVIDER, OLLAMA_URL, OLLAMA_EMBEDDING_MODEL, OLLAMA_LLM_MODEL, CLAUDE_BIN, CLAUDE_MODEL. Cargo: adds ureq (rustls, no extra runtime).

Example — fully local embeddings + Opus reasoning, no keys:

# ~/.patcha/.env
EMBEDDING_PROVIDER=ollama
OLLAMA_EMBEDDING_MODEL=nomic-embed-text   # 768-dim, matches the store
LLM_PROVIDER=claude
CLAUDE_MODEL=opus

Verification (real binary, no cloud login)

• cargo build clean; cargo test passes (incl. sqlite-vec roundtrip).
• Ollama embeddings: collect → 4 events embedded → stored; search returns the correct top match for each query (redis/jwt/ffmpeg/kubectl).
• LLM via Ollama (llama3.2) and via Claude headless (claude -p, Opus): summarize generates a coherent daily summary through both. The Opus path produced an honest summary (declined to invent missing detail) where the small local model hallucinated.

Notes / scope

• Dimension is fixed at 768 in db/migrations.rs, so Ollama embedding models must be 768-dim (nomic-embed-text fits; clear error otherwise).
• The claude backend spawns one claude -p subprocess per LLM call — fine for nightly summarize/compaction; for per-event categorization you'd likely keep LLM_PROVIDER=ollama.
• Unrelated pre-existing bug (left untouched): find_git_repos skips every dotfile before its .git check, so the git collector never discovers repos.
• Implemented with AI assistance (Claude Code).