fix: dataset_load Path P now returns the h5ad; re-pin biodata-registry 0.1.16

feat(de): surface the results-CSV schema; prompt bans schema-sniff re-reads

feat(datasets): dataset_load — execute the loading plan by dataset_id in one call

feat(de): nested-batch guard now recommends the de-confounded alternative

feat(de): default low-count gene filter pre-DESeq2 + stat-ranking prompt rule

feat(de): sample-level library-size pre-flight in decoupler_differential_expression

fix(enrichment): drop non-finite DE stat genes before ulm; report n_dropped_nonfinite

fix(de): apply subset_query on the DESeq2 branch too

feat(mouse): thin mouse-intake path — staging script, ENSMUSG→MGI map, nested-batch guard

fix(tools): keep top_table inside the executor's 4000-char output cap

perf(tools): return the reporting table inline from every chain-terminating tool

fix(reporting): return the top table inline so the report needs no re-read

fix(planner): dataset_plan_analysis must say how to LOAD the dataset

fix(agent,ui,prompt): cut the measured step waste, and stop mislabelling timeouts

fix(loaders): authenticate the h5ad fetch in decoupler_load_url_counts

fix(scoring): resolve gene symbols from var; never let a sample filter no-op

fix(sc,ui): filter lowly-expressed genes after pseudobulk; sweep every output dir

feat(sc): add decoupler_pseudobulk_aggregate — the missing Path P step

fix(scoring): close the six PROGENy per-sample gaps from the GSE205154 eval

refactor: move writable-dir constants to core.paths, fix workflow->tools layer violation

style: apply ruff lint --fix + ruff format across the tree

fix(rna_sc): harden load/visualize UMAP against missing embedding/leiden (#1)

feat(loveless): integrate biodata-registry 0.1.8 sc datasets (Path P by modality)

feat(loveless): sc loaders resolve hosted/private h5ad URLs (ADR-0006)

feat(loveless): ADR-0006 #6 — h5ad source type + Path P loading plan

feat(loveless): ADR-0006 + agent-side sc serving scaffold

perf(mcp): HTTP transport core — add_mcp_http on agent/tool_manager + adata cache + instrumentation

fix(loader): authenticate h5ad loads in dataset metadata tools

feat(integration): ComBat per-sample scoring tool decoupler_pool_cohorts (ADR-0001 item 10)

refactor(loader): centralize authenticated URL->local resolution

fix(loader): authenticate huggingface.co downloads via HF_TOKEN

ADR-0001 Phase 2 Mode A: early integration (T7 + T8 + T9)

feat(concordance): same-cohort normalization concordance routine (decoupler_normalization_concordance)

fix(loaders): accept http(s)/ftp URLs in bulk dataset loaders

feat(sanity): ADR-0002 Phase 2 — auto-fire tissue-identity contamination + PROGENy/Hallmark parity

feat(sanity): ADR-0002 Layer-2 result-aware sanity layer (Phase 1 + S1)

feat: hide microarray from advertised inventory + plain-language Approach in solutions

ADR-0001 Phase 1, steps 5-6 (T5): decoupler_meta_analyze + A->B->refuse wiring

ADR-0001 Phase 1, step 3 (T4): dataset_get_integration_plan wrapper + 0.1.2 re-pin

ADR-0001 Phase 1, step 1: declare Mode-B combine conformance on rna tools

Split bulk_rnaseq.py (1073) into a package preserving the public API

Split dataset_tools.py (1709) into a package preserving the public API

Suppress star-import lint noise in rna package (file-level noqa)

Split rna.py (1847 lines) into an rna/ package preserving the public API

Lint cleanup: remove unused imports, empty f-strings, split import lines

Refactor decoupler_join_clinical_metadata: extract helpers (CC 36->15)

Refactor dataset_count_metadata_values: extract helpers (CC 38->14)

Code-quality fixes: dead assignments, redundant imports, MD5 intent, type hints

Add probe-collapse precompute Phase 2 (collapsed_url support)

Fix 4 GPL-annotation fallback edge cases + 2 integration test bugs