[Grohle]

Summary

Follow-up to #1: the inference side of the read-only copilot, which the UI, Tauri commands and CLI merged in #1 already call.

• crates/assistant/src/backend/candle.rs: pure-Rust GGUF inference with Candle (qwen2 and llama families, CPU), greedy decoding so answers are reproducible, repeat penalty, stop tokens, catch_unwind so a corrupt file is an error and never a panic.
• model.rs: streaming download of the default model (Qwen2.5-0.5B-Instruct Q4_K_M, Apache-2.0, 0.49 GB) with SHA-256 verification and progress events; a .part file is never left as installed. Real hashes pinned.
• tasks.rs + prompts/*.txt: prompts as data (system rules, JSON schema per task, few-shot examples), JSON priming, one retry with the parser error, then a plain-text fallback marked unstructured. Filter values are validated against the current report (unknown values are dropped and listed); YAML suggestions are validated by the real config parser.
• convert.rs: documentation → knowledge.yaml for Markdown, text, YAML, JSON, Word (.docx) and PDF; broken files are reported, never fatal.
• .cargo/config.toml: target-cpu=x86-64-v3 on x86-64 targets so Candle compiles its AVX2/FMA kernels (3–4× faster inference). Documented in the README requirements.
• Clippy fixes CI flagged on VeriModel 1.0.0: IFC QA gate, QTO, BCF export, desktop app, CLI and local assistant #1 (manual_range_contains, large_enum_variant), RAM figure corrected, release notes updated.

Real-model results (ignored test, run locally with the downloaded model)

Task	Result
Ask "What does rule ID-001 check?"	"Every element must have a unique GlobalId." with sources
Filter "errors on Level 1 for walls"	severity: [ERROR], storey: Level 1, ifc_type: IfcWall, dropped: []
Explain a spatial-containment finding	names the element, GlobalId and the fix
Config help "require FireRating on walls"	valid required_psets YAML, identical on re-run

Load 1.5 s, prompt 33–38 tok/s, generation 12–15 tok/s on a 4-core CPU; about 30–50 s per question, 1.0 GB resident.

Test plan

• cargo fmt --all -- --check clean
• cargo clippy --workspace --all-targets -- -D warnings clean
• cargo test --workspace: 168 passed, 0 failed, 2 ignored (real-model smoke test and demo writer, both opt-in)
• Real-model smoke test passes all four checks (cargo test -p verimodel-assistant --release --test real_model -- --ignored)
• Windows installer and CLI cross-compiled with cargo-xwin

---

Generated by Claude Code