feat: Go eval CLI with CGo FFI bridge

[urmzd]

Summary

• Adds a Go eval CLI (apps/eval-cli) that benchmarks base vs GAP flows across multiple providers (Google, OpenAI, Ollama, Groq, GitHub Models)
• Introduces C FFI bindings (src/cffi.rs) so the Go CLI can call the Rust GAP apply engine via CGo
• Fixes Ollama provider ignoring --host flag when OLLAMA_API_KEY is set (was silently redirecting to ollama.com)
• Fixes silent error swallowing from saige SDK ErrorDelta — stream errors are now extracted and surfaced in all runner paths
• Adds justfile recipes: build-go, test-go, run-go, report-go
• Includes experiment 026 results (gemma4 via Ollama): base flow succeeds, GAP envelope apply fails (model capability limitation)

Test plan

• go vet ./... passes
• go test ./... passes
• Smoke tested direct OpenAI adapter vs provider.Build — both produce output
• End-to-end run of experiment 026 with Ollama (gemma4) — base flow produces valid artifacts, GAP flow correctly reports parse/apply failures
• Test with a cloud provider (Google/OpenAI) for GAP apply success validation