[scbrown]

Creating an ONNX session with some system runtimes emits hundreds of duplicate-schema errors despite returning success. Bobbin now validates native session creation in a bounded child process and rejects that runtime with one concise error. Compatible runtimes preserve semantic results; keyword search remains available. The check never downloads libraries or activates CUDA.

Default logging now uses WARN and honors RUST_LOG, so dependency INFO logs no longer print absolute index paths. Runtime discovery notices also omit filesystem paths.

Validation: compile check and executable build passed; six real CLI controls passed, including semantic search/context with a compatible runtime, rejection of the noisy runtime, keyword independence, and explicit logging overrides. HOME guard clean. Focused process controls cover successful-but-noisy children, failed children, diagnostic overflow and timeout/reaping. Hosted CI and independent review remain required. This draft changes source only; no release or installation.

Demo (readable transcript; candidate only, not installed): the same public Flask fixture returned the expected search/context with zero stderr using the compatible runtime. With the affected system runtime, the old installed 0.29.0 emitted 629 schema errors despite exit 0; this source candidate exits 1 with the complete diagnostic below. This is a behavior comparison across installed 0.29.0 and the proposed source, not a claim that the current PR base binary was rebuilt. The private before log contains absolute paths and is not published. The final shared-harness animation waits for the reviewed source landing.

Error: Failed to load embedding model

Caused by:
    ONNX runtime rejected: duplicate schema registration; use a compatible bundled runtime via ORT_DYLIB_PATH

Reproduce with a prepared local index: bobbin search --mode semantic --limit 2 "show the newest blog posts", selecting the affected or compatible runtime through ORT_DYLIB_PATH. BOBBIN_GPU=0 keeps this CPU-only.