neural: real-MNIST external-validity tier (collapse + grounding)

Confirms model collapse and its arrest by grounding on REAL images, not
just the synthetic sandbox. A conv VAE (the canonical generative-collapse
model) is retrained each generation on its own generated digits, with a
fraction g of fresh real MNIST mixed in. Modes = digit class x stroke-
thickness bin (K=30, Zipf, ~18 tail modes); the oracle is a frozen CNN +
deterministic thickness at 98.5% mode accuracy (30x30 confusion matrix
recorded in the manifest as the measurement-noise floor).

Result (4 reps): dry (g=0) collapses to a single mode -- forward-KL
0.5->18, support 30->1, tail 1.0->0.06, H->0 -- while 10% grounding holds
all 30 modes (KL~0.6, full tail, H~0.9). Signs, not magnitudes (blueprint
3.5); the exact synthetic oracle stays the quantitative anchor. The VAE
needs ~10% grounding vs the synthetic histogram's ~5%, consistent with the
grounding finding that trained nets need more than the exact operator.

Plugs into the existing data-agnostic contract (metrics/grounding/output
reused verbatim): mnist_data (thickness bins, class x thickness bijection,
MnistSampler), mnist_oracle (ClassifierOracle + confusion matrix),
mnist_vae (ConvVAEGenerator), mnist_loop (run_mnist_lineage), kind=
mnist_lineage dispatch, MnistCfg/OracleCfg. Figures: plot_mnist (parquet-
only) + mnist_montage (eyeball diagnostic showing digits degenerate to one
blurry mode). make mnist / make env-mnist, kept out of the make neural
loop. 99 tests green (+5 torchvision-gated).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

configs/neural/mnist_collapse.yaml

@@ -0,0 +1,59 @@
+experiment: mnist_collapse
+kind: mnist_lineage
+seed: 20260705
+n_replicates: 4
+
+# (Layer 1.5 external-validity tier; maps to Layer-1 E1/E2 and neural collapse/grounding): does
+# model collapse — and its rescue by grounding — appear on REAL MNIST images, not just the
+# synthetic sandbox? A convolutional VAE (the canonical model in which generative collapse was
+# first observed) is retrained each generation on the previous VAE's own generated images, plus a
+# fraction g of fresh REAL MNIST images (grounding). Modes = digit class x stroke-thickness bin
+# (K=30) resampled to a Zipf p*; a frozen CNN + deterministic thickness is the oracle (its
+# confusion matrix, recorded in the manifest, is the measurement-noise floor). Expect (per E1/E2):
+# the dry arm (g=0) collapses — rare modes die, forward-KL climbs, support shrinks — while a
+# grounded arm holds the tail. This is confirmation-only: SIGNS, not magnitudes (blueprint 3.5);
+# the exact synthetic oracle remains the anchor for every quantitative comparison. Falsifier: the
+# dry VAE shows no diversity loss, or grounding fails to arrest it.
+
+mnist:
+  K: 30
+  n_classes: 10
+  style_bins: 3            # K = 10 classes x 3 stroke-thickness bins
+  R: 1
+  tail: zipf
+  zipf_s: 1.5             # the rarest ~18/30 modes form a real tail (~9% of the mass)
+  tail_threshold: 1.0e-2
+  init: truth
+  data_root: data
+
+model:
+  kind: convvae
+  latent: 32
+  epochs: 30
+  lr: 1.0e-3
+  batch_size: 256
+  beta: 1.0
+  n_eval: 10000          # generate-and-classify samples for the mode-distribution readout
+
+oracle:
+  epochs: 5              # frozen digit CNN (~98.5% mode accuracy = the noise floor)
+  lr: 1.0e-3
+  batch_size: 256
+  cache: models/mnist_cnn.pt
+
+dynamics:
+  n: 6000                # images the pupil VAE sees per generation (drift strength)
+  grounding: {m: 0, policy: proportional}   # m overwritten per g by the sweep
+
+generations: 15
+
+metrics:
+  kl_floor: 1.0e-9
+  support_eps: 1.0e-9
+
+sweep:
+  - param: g
+    values: [0.0, 0.1]   # dry vs grounded (VAE collapse is strong; needs ~10% real, cf. grounding)
+
+output:
+  dir: results/mnist_collapse