Group examples by data type and fix the index

[afloresep]

Follow-up to #39, which is unrelated in content — this branch is off master and touches no source code.

examples/ had grown to 19 files in one flat directory. Two things were wrong with it.

The index was mostly fiction

examples/README.md listed four examples that were never committed — sar_egfr.py, afdb_clusters_tmap.py, singlecell_trajectory_tmap.py, singlecell_reprogramming_tmap.py. There are no deletion records in git history; they never existed. Meanwhile it omitted eleven files that do exist, including every image example except two.

So the index advertised 8 examples, half of them imaginary, and hid most of the real ones. The README is rewritten and now lists all thirteen with what each actually does.

New layout

examples/
  README.md
  cluster_65053.csv
  chemistry/   molecules_tmap.py  smiles_tmap.py  molecules_tmap_legacy.py
  images/      pet_breed_audit.py  mnist_cosine_tmap.py  emnist_characters_tmap.py
               flowers_tmap.py  cub200_birds_tmap.py  wikiart_tmap.py
  proteins/    esm_atlas_tmap.py
  text/        word_embeddings_tmap.py  word_embeddings_50k.py
  layout/      untangle_demo.py
  data/        (gitignored download cache, unchanged)

cluster_65053.csv deliberately stays at the root. It is referenced from 3 docs, 4 scripts and 5 notebooks — moving it would churn 13 files including notebook JSON, and break any snippet a user has already copied out of the docs.

Removed: 10 MB and a tuning experiment

adaptive_coef_exp_3panel.py is an 8×8 sweep of layout coefficients across 2k/64k/200k molecules with a --probe flag to "time risky 200k corners". That's how the defaults were chosen, not something a user runs.

Its only dataset, chembl_200k.csv (10 MB), was used by nothing else in the repository. Both are gone, which takes roughly 72% of the tracked weight of examples/ out of every clone. Still in git history if it's ever wanted back.

untangle_demo.py is kept, under layout/ — it's the only documentation of what UntangleMode does.

Paths

Every script anchored its data, cache and output at Path(__file__).parent, which breaks when the script moves one level deeper — including against the gitignored examples/data/ cache that may already hold gigabytes of downloads. Each moved script now anchors at

EXAMPLES_DIR = Path(__file__).resolve().parents[1]

so the shared CSV and the existing cache keep resolving, and outputs still land in examples/. Usage lines in the docstrings are updated to the new paths.

mnist_cosine_tmap.py was writing to a cwd-relative "examples/mnist_tmap.html", so it only worked from the repository root. It uses the same anchor now.

Also

• cazyme_analysis.ipynb moves to notebooks/14_cazyme_analysis.ipynb, next to the other thirteen. The root README's notebook table gains rows 13 and 14 — 13 was already missing.
• Four docs updated for the moved script paths (docs/molecule_tutorial.md, lshforest_guide.md, layout_guide.md, minhash_guide.md).

Checked

• Ran examples/chemistry/molecules_tmap.py from /tmp: it found the CSV and wrote output back to examples/, confirming the anchoring is cwd-independent.
• All 13 scripts parse; every README link resolves; every script appears in the README.
• 454 passed, 33 skipped. ruff check src/ tests/ clean, which is what CI gates.
• ruff check examples/ goes from 26 errors to 20. The remaining 20 are pre-existing in esm_atlas_tmap.py (long lines, trailing whitespace) and left alone — examples/ is not linted by CI, and fixing them would bury the move in unrelated churn.