docs: state the working layout and the pseudo-label taxonomy

The reproduce commands mix paths from two repositories without saying so; show the
directory they assume and how to fetch a cohort without cloning 1.05 TB.

bach_cellvit CSVs carry PanNuke codes 1-5, which no file explained, and they are a
different taxonomy from the 4-class NuCLS-super one behind tail_metrics/.

DATA.md

@@ -268,11 +268,16 @@ below 1 % everywhere, so this is not a handful of pathological graphs.
 macro-mean, which runs 0.75–0.82 on the labelled cohorts.
 
 Contents: `tiles/` holds 109,904 224 px tile graphs with CellViT++ pseudo-labels (BACH 13,837,
-BRACS 95,670, BreakHis 397); `bach_cellvit/` holds 400 per-image pseudo-label CSVs (538,421
+BRACS 95,670, BreakHis 397); `bach_cellvit/pseudolabels/` holds 400 per-image pseudo-label CSVs (538,421
 cells, columns `centroid_x_px,centroid_y_px,cell_type,type_prob`) plus a summary JSON; and
 `paper_metrics/` holds the pooled edge-homophily and pseudo-label-gate results as JSON, plus
 `tail_metrics/<cohort>.json`, which is the artifact behind the table above: `nodes` is
 `n_nodes_total - n_isolated_total` and the two tail columns are the per-node fractions below
+`cell_type` in those CSVs uses PanNuke's codes, with background dropped at inference:
+1 = Neoplastic, 2 = Inflammatory, 3 = Connective, 4 = Dead, 5 = Epithelial. Note this is a
+different taxonomy from the 4-class one behind `tail_metrics/` and `pseudolabel_validation.json`,
+which come from the NuCLS-super head; the two are separate analyses.
+
 0.3 and 0.5. Do not read `bach_cellvit/bach_fattail_cellvit.json` as the BACH row of that
 table; it is a different analysis (whole-image CellViT graphs, 538,421 cells) rather than the
 224 px tile graphs (439,533 nodes) the table is computed over.
@@ -374,6 +379,23 @@ The training, evaluation and analysis scripts referenced below live in
 [github.com/Ace3Z/GrapHist-V2](https://github.com/Ace3Z/GrapHist-V2). This repository ships
 the data they consume, plus the model and its loader.
 
+The commands assume both are in one working directory, the code repository with this
+repository's contents beside it:
+
+```
+GrapHist-V2/            <- git clone of the code repository
+├── src/  studies/  experiments/  modeling/ ...
+├── graphs/            <- from this repository
+├── embeddings/
+├── labels/
+└── baselines/
+```
+
+So `python modeling/rebase_graph_paths.py` runs the copy that ships here, while
+`python src/train/generate_embs.py` runs the one from the code repository. Fetch what you need
+with `hf_hub_download`, or `huggingface-cli download Ace3Z/graphist-v2 --include "graphs/bach/*"`
+for a whole cohort; the repository is ~1.05 TB, so do not clone it.
+
 Run the `rebase_graph_paths.py` step first in every case.
 
 **Slide-level MIL subtyping**