feat(api): dispatch /predict/mri via MRI_MODEL_KIND env var

feat(models): selector dispatch for volumetric vs 2D MRI models

feat(models): add 2D resnet18 4-class Alzheimer's MRI inference module

fix(fusion): correct logit-scale comment; top_disease=None on empty input

test(fusion): end-to-end smoke + README section

feat(agents): register run_fusion tool for multi-modal disease confidence

feat(api): add POST /fusion/predict route for multi-modal fusion

feat(fusion): add core multi-modal fuse() with per-disease attribution

feat(fusion): map modality predictions to per-disease signals

feat(fusion): add clinical-test signal normalisers (MMSE/MoCA/UPDRS/gait/age)

feat(fusion): add disease/modality weight registry

fix(fusion): drop unused protected_namespaces; pin tighter bound tests

feat(fusion): add pydantic data contract for multi-modal fusion

test(agents/orchestrator): pin caplog handler to INFO (hardens against pytest variance)

fix(agents/orchestrator): log dropped out-of-stage tool calls (was silent)

test(agents/routing): assert sites_csv key on NIfTI route (depth parity with EEG test)

test(agents/routing): regression tests for MRI/EEG/BBB routing heuristic

test(mri/model): drop redundant caplog.at_level (direct handler attach is the actual mechanism)

fix(mri/model): warn when label_names length != model output dim (was silent override)

Add project files

fix(agents/live): skip on NotFoundError (model availability churns on OpenRouter free-tier)

fix(agents/live): also gate live test on BBB model artifact (avoids dotenv-loaded key path)

feat(api): POST /agent/run endpoint (orchestrator + RAG, stub-injectable)

fix(agents/tools): parameterize processed_dir + translate HTTPException → ValueError

feat(agents): orchestrator loop (function-calling + tool trace + max-steps gate)

feat(agents): Tool dataclass + registry + 4 tool wrappers (3 pipelines + RAG)

feat(rag): RAGRetriever (load + search → chunks with scores)

feat(rag): ingest CLI (markdown/PDF → chunks → FAISS) + sample KB fixtures

fix(rag/store): copy vectors before in-place normalize_L2 (no caller mutation)

feat(rag): FAISS inner-product store with chunk metadata + roundtrip

feat(rag): fastembed wrapper (Embedder, bge-small-en-v1.5, 384-dim)

feat(rag): paragraph-aware chunker (chunk_text)

fix(test): attach caplog handler to non-propagating explainer logger

test(llm): tighten live integration test (slow marker, log-gated skip, on-topic assertion)

test(llm): add network-gated end-to-end OpenRouter integration test

test(llm): pin 401 short-circuit + 400 try-next-model behavior (red)

test(deploy): assert Dockerfile and Dockerfile.hf stay byte-identical

test(deploy): scan ENV lines only in MLflow kill-switch guard (avoid comment false-positive)

test(deploy): assert Dockerfile seeds EEG fixture, runs all pipelines, drops MLflow kill-switch (RED)

fix(frontend): EEG default path points at fixture (HF container has no data/raw/eeg.fif)

feat(deploy): Hugging Face Spaces Dockerfile + supervisord launcher

feat(api): GET /experiments/runs + POST /experiments/diff (Track 5)

feat(api): POST /explain/eeg + /explain/mri — full-stack Track-1 coverage

feat(llm): modality dispatch — explain(payload, modality) for BBB/EEG/MRI

feat(api): POST /explain/bbb — natural-language rationale endpoint

feat(llm): explainer with deterministic template + OpenRouter fallback

feat(api+frontend): MLflow provenance badge in decision card

feat(api): drift z-score in /predict/bbb response

feat(models): train-time confidence stats stashed on _neurobridge_train_stats

feat(api): POST /pipeline/mri/diagnostics — pre/post ComBat KPIs + long-format rows