sync with main 76bcc66: card.json (references / card text)

card.json

@@ -55,7 +55,7 @@
   "size_gb": 0.23,
   "horizon": "full",
   "eval_times": "native",
-  "overview": "## The problem\n\nSet a slender elastic bar moving along its axis and stop it dead at one end:\na compression front leaps from the impact end, sweeps back along the bar at\nthe material wave speed, reflects off the free end, and keeps traversing \u2014\nthe textbook one-dimensional stress-wave problem. The particles barely move;\nthe travelling stress wave *is* the physics. That leaves a learned surrogate\nnothing to hide behind: it must carry a sharp front through hundreds of\nautoregressive steps and keep the wave speed, the reflection timing, and the\nstress amplitude honest over 4\u201311 bar traversals, depending on length.\n\nStructBench ships the 2D SPH version: an LS-DYNA `*MAT_ELASTIC` strip\n(scaled toy constants) five particle rows deep, 8 mm wide and 200 / 300 /\n400 / 500 mm long, arrested from an initial axial velocity of 1\u20138 m/s\n(wave speed ~70.7 m/s). The task is an **autoregressive next-step\nsurrogate** \u2014 from a short ground-truth prefix the model advances the\nparticle state one output step at a time over the 30 ms record, predicting\nposition and the per-particle axial stress. Axial stress is the headline\ntarget and position the sanity check \u2014 the reverse of Taylor 2D's emphasis.\n\n![Schematic of the wave propagation setup: a strip of elastic particles moving axially toward a fixed boundary at one end.](../../assets/problem_wave_propagation.png)\n\n*Problem setup: an elastic bar strip (length L = 200\u2013500 mm) moving axially\nat 1\u20138 m/s is arrested dead at one end.*\n\n## The entry tier\n\nThis is the platform's entry benchmark \u2014 16 linear-elastic cases, 0.23 GB \u2014\nsized for onboarding, the docs tutorial, and CI-scale runs rather than for\nseparating strong methods. The split is an interior holdout on the 4-length\n\u00d7 4-velocity grid, interpolation only: every length and every velocity\nappears in training, `val` (300 mm at 2, 400 mm at 4 m/s) only picks each\nrun's checkpoint, and `test_interp` (300 mm at 4, 400 mm at 2 m/s) is\nscored. Everything is reported in physical units \u2014 axial-stress RMSE in MPa,\nposition RMSE in mm \u2014 and the quantities of interest read the wave directly:\narrival time at the 25 / 50 / 75 % gauge stations and the peak stress. The\nreference CGN baseline nails the arrival times and overshoots the peak; the\nnumbers are below.",
+  "overview": "## The problem\n\nSet a slender elastic bar moving along its axis and stop it dead at one end:\na compression front leaps from the impact end, sweeps back along the bar at\nthe material wave speed, reflects off the free end, and keeps traversing \u2014\nthe textbook one-dimensional stress-wave problem. The particles barely move;\nthe travelling stress wave *is* the physics. That leaves a learned surrogate\nnothing to hide behind: it must carry a sharp front through hundreds of\nautoregressive steps and keep the wave speed, the reflection timing, and the\nstress amplitude honest over 4\u201311 bar traversals, depending on length.\n\nStructBench ships the 2D SPH version: an LS-DYNA `*MAT_ELASTIC` strip\n(scaled toy constants) five particle rows deep, 8 mm wide and 200 / 300 /\n400 / 500 mm long, arrested from an initial axial velocity of 1\u20138 m/s\n(wave speed ~70.7 m/s). The task is an **autoregressive next-step\nsurrogate** \u2014 from a short ground-truth prefix the model advances the\nparticle state one output step at a time over the 30 ms record, predicting\nposition and the per-particle axial stress. Axial stress is the headline\ntarget and position the sanity check \u2014 the reverse of Taylor 2D's emphasis.\n\n![Schematic of the wave propagation setup: a strip of elastic particles moving axially toward a fixed boundary at one end.](../../assets/problem_wave_propagation.png)\n\n*Problem setup: an elastic bar strip (length L = 200\u2013500 mm) moving axially\nat 1\u20138 m/s is arrested dead at one end.*\n\n## The entry tier\n\nThis is the platform's entry benchmark \u2014 16 linear-elastic cases, 0.23 GB \u2014\nsized for onboarding, the docs tutorial, and CI-scale runs rather than for\nseparating strong methods. The split is an interior holdout on the 4-length\n\u00d7 4-velocity grid, interpolation only: every length and every velocity\nappears in training, `val` (300 mm at 2, 400 mm at 4 m/s) only picks each\nrun's checkpoint, and `test_interp` (300 mm at 4, 400 mm at 2 m/s) is\nscored. Everything is reported in physical units \u2014 axial-stress RMSE in MPa,\nposition RMSE in mm \u2014 and the quantities of interest read the wave directly:\narrival time at the 25 / 50 / 75 % gauge stations and the peak stress. The\nreference CGN baseline nails the arrival times and overshoots the peak.\nDespite its entry-tier size, the benchmark separates prediction schemes\nsharply (2026-08-28 multi-method fleet): the time-conditioned operators\n(Transolver, GeoFLARE) beat CGN ~3x on the fields and ~10x on peak stress,\nwhile every mesh-native autoregressive family (MGN, Transolver, GeoFLARE)\nfails outright (relative L2 > 1) - the sustained 30 ms reverberation\nrecirculates rollout errors instead of letting them decay, and only CGN's\nrelative-displacement particle-graph formulation survives it. The\nautoregressive rows are kept as deliberate negative results; the numbers\nare below.",
   "figures": [
     {
       "path": "assets/wave_rollout.gif",