Upload NIGHT6_STATE.md with huggingface_hub

NIGHT6_STATE.md

@@ -33,6 +33,21 @@ covariate regression so "length effect" isn't just "more rare words". Report N p
 - JSON -> local night6_json/ + HF. Figs -> figs_night6/. Secrets only from ~/secrets.env.
 
 ## LOG (newest at top)
+- 03:45 — NIGHT COMPLETE. L3 proc finished all compute but crashed on final write (night6_json/ dir missing on box,
+          FileNotFoundError) — RECOVERED all numbers from L3_full.log, reconstructed L3_recomb_models.json locally + pushed HF.
+          Uniform m128 row recovered (.5534/.4046/.5682/.8052/1.0923); decode-SBERT all 4 bin/width combos confirm attn≈uniform
+          (Δ -.006..+.008) & DeepSets worse => no meaning-transfer. §20 numbers all verified consistent. FINAL SWEEP: marble/photo/lyra
+          all who=0, GPU 0MiB, zero stray mine-procs. Deliverables on HF: INTERP_RESULTS.md §20 (3536L) + fig_length_stratification.png
+          + L1/L1b/L2/L2b/L3 JSONs. Memory (capstone + index) updated. Boxes idle. WIND DOWN.
+- 03:20 — §20 length-stratification chapter WRITTEN into INTERP_RESULTS.md (3536 lines) + fig_length_stratification.png
+          (4-panel) + pushed to HF. Decode-SBERT secondary confirms L3 (m128 long: uniform .301/attn .300/deepsets .204,
+          Δ=-.001 => no meaning-transfer). L3 JSON still finalizing (decode tail slow); scheduled wakeup to collect + final sweep.
+          HEADLINE §20: length = a CAPACITY/dilution axis, NOT a structure axis. (1) recon degrades gracefully + monotonic,
+          genuine 1/N dilution (89% survives rare-word match), field-meaning robust (SBERT .81 @43tok). (2) content readouts =
+          single length-INVARIANT directions (one global probe suffices, natural+templated); only relational probe varies = a
+          short-sentence surface-position leak (cos -.93), sharpens §19 no-binding not overturns. (3) dilution = per-token CAPACITY
+          not aggregation: attn-pool/DeepSets ≈ uniform 1/N at fixed bottleneck, learned weights don't favor content. OWED: L4
+          cross-encoder length port; natural relational gold label.
 - 03:07 — L3 (photo) near-done. RESULT: at fixed-capacity bottleneck FVU RISES w/ length (m128 short .41->vlong 1.09;
           m64 short .48->vlong 1.23) => dilution real once pooling lossy. BUT attention-pool ≈ uniform 1/N (m64 attn .47/.65/.92/1.22
           vs uniform .48/.66/.92/1.23); DeepSets WORSE. Learned attn weights: NO content-upweight (Spearman a·c_i_norm=-0.09),