docs: public-readiness audit fixes (stale refs + genericize Euler mentions)

README.md

@@ -67,8 +67,9 @@ The skill runs as a guided, three-stage workflow (the detailed algorithm lives i
 2. **Plan.** Explore the object, gather prior knowledge for the tissue/context from the
    knowledge stack, draft the label hierarchy and per-level vocabulary, present for review.
 3. **Execute (top-down, recursive).** For each level, and within each parent group:
-   assemble candidate types + markers, panel-gate them, score (Wilcoxon + COSG markers;
-   AUCell/pyUCell signatures with a z-scored + consolidated cross-check), cross-check
+   assemble candidate types + markers, panel-gate them, score (Wilcoxon DE + a
+   `filter_rank_genes_groups` specificity pass; AUCell/pyUCell signatures with a
+   z-scored + consolidated cross-check), cross-check
    against a tissue-matched classifier / a reference mapping (CellMapper) as a
    *hypothesis*, sign off the assumptions ledger, profile ambiguous clusters before any
    `Unknown`, write per-cluster evidence cards, then assign a label — or sub-cluster and
@@ -143,7 +144,7 @@ skills/cell-type-annotation/
 │   ├── artifacts.md               # working-dir layout, run_state schema, reports
 │   ├── compute_environments.md    # the 3 compute profiles, env detection, scheduler handoff
 │   ├── tool_registry.md           # canonical step → cta command → GPU/profile map
-│   ├── marker_tools.md            # decoupler, pyUCell, COSG, snapseed
+│   ├── marker_tools.md            # decoupler, pyUCell, snapseed
 │   ├── automated_tools.md         # CellTypist, CellMapper, scANVI/scArches, Census, what to skip
 │   ├── knowledge_sources.md       # MCP coverage, CL normalization, atlases, loaders
 │   └── differential_expression.md # markers (Job A) vs condition DE (Job B)

skills/cell-type-annotation/SKILL.md

@@ -111,7 +111,7 @@ and `plan.md` in the working directory first, then continue from
 - Verify whichever MCPs are present actually respond before proceeding.
 
 **1b. Detect the compute environment (do this early — it shapes the plan).** Run
-`cta detect-env` (stdlib-only — works with any python). It reports the accelerator
+`cta detect-env` (no heavy imports — no scanpy/torch, so it's a fast cold start). It reports the accelerator
 (CUDA / Apple-Silicon MPS / none), usable RAM (clamped to cgroup/scheduler limits), any
 batch scheduler, and a **suggested profile**: `local_cpu` / `local_gpu` / `scheduler`.
 **Confirm it with the user** — and you *must* ask in the one ambiguous case the probe flags:
@@ -383,15 +383,15 @@ first, so a deep recursion never leaves the user lost.
    an L1 compartment (NPC vs Neuron), *qualify the label by compartment*
    (`Ventral telencephalic NPC` vs `… neuron`) — never reuse one label across two parents.
    Verify any time with `cta report write-back --dry-run` (runs
-   `cta_io.check_hierarchy_nesting`, refusing a non-nested hierarchy). Why the CSV is the
+   `cta.io.check_hierarchy_nesting`, refusing a non-nested hierarchy). Why the CSV is the
    source of truth, and the full card schema: `reference/methodology.md` §4.
 7. **Critic pass.** Spawn independent critic subagents (Agent tool) that read the
    exported **raw tables** (not just the card), re-derive markers, fetch ≥1 external
    hypothesis, and try to **refute** each label — biologically **and hierarchically**:
    is the label a coherent refinement of its parent path (flag maturation/regional
    mismatches like an "IPC"/"neuron" L3 under an "NPC" L1 — the kind the deterministic
    nesting check can't catch)? Reconcile; revise or downgrade. (Structural nesting itself
-   is enforced in code — `cta_io.check_hierarchy_nesting` / `cta report write-back`.)
+   is enforced in code — `cta.io.check_hierarchy_nesting` / `cta report write-back`.)
 8. **Report, gate & descend.** Build the level's HTML report with
    `cta report build` (pass `--adata --cluster-key --viz-key`, plus
    `--embedding-key` for the similarity-ordered palette, to embed the numbered
@@ -465,7 +465,7 @@ and **whether it needs a GPU**, `reference/tool_registry.md` is the canonical ma
 | `cta reference train-model` | GPU/large-memory: train (or refine) a scANVI reference model on a labeled atlas (scvi-tools). Reference side of the train-then-map workflow. |
 | `cta reference map-scarches` | GPU/large-memory: map the query onto a scANVI reference (a model you trained, or a shipped scvi-tools hub model) via scArches surgery — labels + per-cell uncertainty, in the `refmap_*` schema. |
 | `cta crosscheck scimilarity` | Optional **HUMAN-only** cross-check: zero-shot per-cell labels via SCimilarity (~16–27 GB RAM, CPU, ~9 GB annotation bundle). Foundation model → additional hypothesis only, never primary; most useful on broad/multi-tissue human queries where no CellTypist/SingleR model fits. Needs the model bundle — provision it once with `cta reference fetch-scimilarity`. |
-| `cta reference fetch-scimilarity` | One-time provisioning for `cta crosscheck scimilarity`: download + extract the SCimilarity model bundle. Annotation-only by default (~9 GB on disk; ~30 GB resumable download, skips the unused 32 GB cellsearch). Point `--model-path` at the resulting `<dest>/model_v1.1`. Needs network (`module load eth_proxy` on a compute node). |
+| `cta reference fetch-scimilarity` | One-time provisioning for `cta crosscheck scimilarity`: download + extract the SCimilarity model bundle. Annotation-only by default (~9 GB on disk; ~30 GB resumable download, skips the unused 32 GB cellsearch). Point `--model-path` at the resulting `<dest>/model_v1.1`. Needs network (ensure outbound internet on a firewalled compute node). |
 | `cta reference fetch-census` | Optional: pull a capped (or, on a large-memory profile, larger) tissue-matched reference subset from CELLxGENE Census (no full-atlas download) to feed `cta crosscheck reference-map` when the user has no atlas and no CellTypist model fits. Needs the `cellxgene-census` extra + network. |
 | `cta crosscheck obs-pred` | Cross-check from an EXISTING in-object prediction column (e.g. a previously-mapped atlas `*_pred`) — no recompute. Prefer over `cta crosscheck reference-map` when such a column exists and fits the level. |
 | `cta cluster subcluster` | Refine one/more clusters: Leiden on the anchor embedding + per-sub marker-fraction / QC / composition summary. Standardized sub-clustering (in scope). |

skills/cell-type-annotation/pixi.toml

@@ -91,7 +91,7 @@ extra-index-urls = ["https://pypi.nvidia.com"]
 
 # scimilarity pulls hnswlib, but its PyPI wheel is the problem: the 0.8.0 cp312 wheel is built
 # with AVX-512 and SIGILLs `cta crosscheck scimilarity` when SCimilarity loads its prebuilt
-# hnswlib kNN index on AVX2-only CPUs (verified crashing on Euler's AMD EPYC 7763 / Intel Xeon
+# hnswlib kNN index on AVX2-only CPUs (verified crashing on AMD EPYC 7763 / Intel Xeon
 # E3-1284L v4 — neither has AVX-512). The fix is to source hnswlib from conda-forge instead:
 # conda-forge builds it portably (baseline microarch, no `-march=native`), so the same 0.8.0
 # runs on any CPU (verified: the conda-forge .so has zero AVX-512/AVX2 ops and the full
@@ -104,7 +104,7 @@ hnswlib = ">=0.8"
 # `rank_genes_groups`) are built against CUDA 13 and link `libcudart.so.13`, but the rest of the
 # GPU stack is CUDA 12 (cuml-cu12 / cupy-cuda12x / torch-cu12, which load libcudart.so.12). Without
 # a CUDA-13 runtime present the kernel .so fails to load, rapids-singlecell sets it to None, and the
-# GPU DE path silently falls back to scanpy (verified on Euler). Ship the CUDA-13 cudart runtime from
+# GPU DE path silently falls back to scanpy (verified in practice). Ship the CUDA-13 cudart runtime from
 # conda-forge so libcudart.so.13 lands in $ENV/lib alongside .so.12 — the two majors coexist (the
 # driver supports both) and the kernel loads. This is the actual enabler of issue #38's GPU path.
 cuda-cudart = "13.*"
@@ -128,7 +128,7 @@ cudf-cu12 = "*"
 # profile (atlas-scale ranking in seconds vs ~minutes on scanpy/CPU). Drop-in — same wilcoxon
 # method, same `uns['rank_genes_groups']` output the CPU path writes; scanpy stays the CPU default.
 # The `rapids-cu12` extra pulls the FULL RAPIDS stack it imports at load (cudf/cuml/cugraph/cuvs);
-# the bare cuml-cu12/cudf-cu12 above (for CellMapper's kNN) are not enough — verified on Euler that
+# the bare cuml-cu12/cudf-cu12 above (for CellMapper's kNN) are not enough — verified in practice that
 # `import rapids_singlecell` needs cuvs too.
 rapids-singlecell = { version = "*", extras = ["rapids-cu12"] }
 # rapids-singlecell 0.15.2 also imports `dask.array` at module load (rapids_singlecell/_compat.py)

skills/cell-type-annotation/reference/compute_environments.md

@@ -108,7 +108,7 @@ CPU host that merely has the client tools installed. `cta detect-env` flags it;
   annotation kNN; the full ~41 GB download incl. the cell-search TileDB is NOT needed). Fits a
   32 GB machine — optional, and HUMAN-only (skip for non-human queries). Provision the bundle
   once with `cta reference fetch-scimilarity --dest DIR` (resumable ~30 GB download, extracts
-  annotation-only by default; on a compute node `module load eth_proxy` first).
+  annotation-only by default; on a firewalled compute node ensure outbound internet first).
 
 ## Which tool at which step (and what it needs)
 

skills/cell-type-annotation/reference/methodology.md

@@ -272,7 +272,7 @@ hierarchy must nest, so the critic also checks each label against its **parent p
 (L1→…→L_k):
 
 - *Structural* nesting (every finer label has exactly one parent) is a deterministic
-  invariant enforced in **code** — `cta_io.check_hierarchy_nesting`, run by
+  invariant enforced in **code** — `cta.io.check_hierarchy_nesting`, run by
   `cta report write-back` (and any time via `--dry-run`). Don't rely on an LLM to never miss it;
   the critic need not re-derive it.
 - *Semantic* path-consistency is the critic's job — flag a path that **passes** the

skills/cell-type-annotation/src/cta/fetch_scimilarity_model.py

@@ -13,8 +13,9 @@
     unless --keep-tarball), plus ~9 GB extracted → ~39 GB peak, settling to ~9 GB.
   * with --with-cellsearch: ~41 GB extracted (only if you also want cell SEARCH).
 
-The download is resumable (re-run the command to continue a partial download). On an Euler
-COMPUTE node remember to ``module load eth_proxy`` first; a LOGIN node has internet directly.
+The download is resumable (re-run the command to continue a partial download). On a firewalled
+compute node, make sure outbound internet is available first (e.g. load your site's proxy
+module); an internet-connected (login) node works directly.
 
 Usage:
   cta reference fetch-scimilarity --dest /path/to/models
@@ -67,8 +68,8 @@ def _download(url: str, tarball: Path) -> None:
     except subprocess.CalledProcessError as e:
         raise SystemExit(
             f"ERROR: download failed (exit {e.returncode}). It is resumable — re-run the "
-            "same command to continue. On an Euler COMPUTE node, `module load eth_proxy` "
-            "first (a login node has internet directly)."
+            "same command to continue. On a firewalled compute node, ensure outbound "
+            "internet/proxy is configured first (an internet-connected node works directly)."
         ) from e