fly-connectome 2026-09  --  code and results behind seven notes on tsurutalab.org

  notes/?id=larval-connectome-recurrent-null   Track A  scripts_a1 / a3 / a2
  notes/?id=fly-hash-embedding-width           Track B  scripts_b1 / b2 / b3
  notes/?id=malecns-scaling-factor             Track C  scripts_c1 / c2
  notes/?id=fly-hash-intrinsic-dimension       Track B  scripts_b4   (corrects the B note)
  notes/?id=fly-model-teaching                 Track D  scripts_d1 / d2

Environment used: Python 3.11.4, torch 2.5.1+cu121, torchvision 0.20.1+cu121,
python-igraph 1.0.0, pyarrow 25.0.1, h5py 3.16, numpy. GPU: RTX 3060 12GB
(CPU works, slower).

Data is not included. Fetch into data/ as follows:
  data/larval_raw/Supplementary-Data-S1/   Supplementary-Data-S1.zip from
                                           github.com/brain-networks/larval-drosophila-connectome
  data/malecns_repo/                       git clone github.com/flyconnectome/2025malecns
                                           (uses supplemental_data/mcns_fw_edge_comp.feather)
  data/emb/glove-100-angular.hdf5          ann-benchmarks.com
  data/emb/coco-t2i-512-angular.hdf5       github.com/fabiocarrara/str-encoders/releases/tag/v0.1.3
  data/mnist/ data/fashion/ data/cifar_grey/   downloaded by torchvision on first run

scripts_b4_synthetic.py needs no data at all: it generates its own.

Run order and the exact commands are in docs/FINDINGS_2026_09_13.md and
docs/FINDINGS_2026_09_13b.md; the teaching experiments in docs/FINDINGS_D_2026_09_13.md.
results/ holds the numbers the notes cite, as written by the scripts.
