[martyn]

Summary

Introduces EMPF (Exponential Moving Prototype Flow) (empf-v1 + tuned empf-v2), a standalone, clock-free, constant-learning-rate ParticleGAN formulation inspired by e-means (Bessa et al., arXiv:2611.09501), registered as a first-class trainer family without altering DEFAULT_RECIPE = "bispf_optimized" or any incumbent recipe hashes.

Core Formulation (empf-v1 + empf-v2 Refinements)

1. Clock-Free Online e-means Sufficient-Statistics Prototype Bank (particlegan/_empf/prototypes.py, critic.py):
   • Tracks running prototype feature sums $S \in \mathbb{R}^{K \times D}$ and effective mass counts $N \in \mathbb{R}^K$ ($C_k = S_k / N_k$) with hit-rate-scaled EMA rate $\alpha_{\mathrm{eff}}(K, B_r) = \operatorname{clip}(\alpha_0 (B_r / K)^\eta, \alpha_{\min}, \alpha_{\max})$ — zero dataset-size $n$ or epoch-horizon $T$ parameters.
   • Deterministic mass-conserving proportional-split dead-prototype revival: when $N_{\mathrm{dead}} < \tau_{\mathrm{rev}} \bar{N}$, splits an active donor selected via golden/plastic ($\phi_1, \phi_2$) quasi-Monte Carlo CDF sampling while conserving exact total mass ((N_donor + N_dead) / 2) with zero RNG consumption.
2. Mass-Weighted Prototype-to-Fake Entropic Sinkhorn Chords & State-Driven Settling EBI (particlegan/_empf/transport.py, penalty.py):
   • Solves scale-free log-domain Sinkhorn with non-uniform prototype row marginals $w_k = N_k / \sum_j N_j$ against uniform generated column marginals $1 / B_f$ in $O(B \cdot K)$ rather than $O(B^2 \cdot D_{\mathrm{img}})$, preserving rare-mode transport mass while aligning conditioned fibers automatically.
   • empf-v2 State-Driven Settling Born-Infeld Sharpness: Uses softer far-from-manifold sharpness $\beta_0 = 6.0$ during coarse mode acquisition ($r_{\mathrm{ema}} \ge 0.16$) and tightens quadratically once normalized transport residual settles: $s_{\mathrm{settle}} = \operatorname{clip}(1 - r_{\mathrm{ema}}/0.16, 0, 1)$, $\beta_{\mathrm{eff}} = \beta_0 + g(B_r),s_{\mathrm{settle}}^2$ with prototype-bank resolution gain $g(B_r) = 6.5 + 15.5\operatorname{clip}((B_r - 64)/64, 0, 1)$.
   • empf-v2 Discrete-Simplex Post-Acquisition Sinkhorn Thermal Cooling: Damps post-acquisition mini-batch jitter ($r_{\mathrm{ema}} < 0.070$, floor 0.35) on unconditioned discrete-simplex couplings (dim >= 170, B_r, B_f >= 64) using purely instance-bound critic state (EMPFStepRecord, zero step clocks).
3. Degree-of-Freedom-Aware Latent Prior Regularizer (particlegan/_empf/prior.py):
   • Combines an initial-moment covariance tether, scale-free $k$-NN log-repulsion, local scatter circularization, a short-range one-sided discrete codebook non-collision hinge ($N \le \mathrm{DoF}$, zero .item() host sync), and a smooth dead-particle revival tether ($d_{1\mathrm{nn}}^2 / (\text{isolation_ratio} \cdot r_{\mathrm{med}}^2) - 1$).

Forge CUDA Readout (empf-v2, Seed 0, NVIDIA RTX A6000)

Executed commit: 49537dfe0349daec307607f340d6379537d874b9 (readout-05eaed46ab0e45214ae8e7b4.json, 0 INCOMPLETE, 1109/1109 attempt coverage; empf-v1 baseline preserved byte-for-byte at cb5e484b059d3474f7259c40beb7acb4b5bb8704 / readout-d46a41a1da2da54d67d154f4.json):

Selected Recipe / Family	Tier 1 (27)	Tier 2 (78)	Tier 3 (10)	Total View Reqs (115)	Full-Suite 61 PASS	Tuning Budget
EMPF (empf-v2, selected)	27/27	55/78	0/10 (3 blocked)	82/115	22	0 searches / 2 cards
BISPF (bispf-optimized-quintic-v1)	22/27 (4 not run)	55/78 (4 not run)	4/10	81/115	18	0 searches / 19 cards
EMPF baseline (empf-v1, retained)	27/27	48/78	0/10	75/115	20	untuned (1 card)
BCAP (recipe-prior-refactor-bcap-pure-v1)	22/27 (4 not run)	49/78 (4 not run)	0/10	71/115	16	85 searches / 473 cards
GAN v3 / K3P / KA2 / R1-R2	20/27 (4 not run)	0/78	0/10	20/115	5	8–46 cards (tuned)
Atlas (atlas-particle-normal-v4)	18/27	0/78	0/10	18/115	8	5 cards
E22 (e22)	0/27	0/78	0/10	0/115	0	untuned (1 card)

Disclosure: BISPF and BCAP pinned snapshots predate the 4 newer VAE/AE Tier 1 tasks (4 not run).

Key Scientific Gains (empf-v1 $\to$ empf-v2, Zero Regressions)

• Declare experiment dependencies and add CPU test CI #1 Overall on Counted View Requirements (82/115) & Fixed Full-Suite Passes (22/61): Retains all 20 empf-v1 passes (11/11 required Tier 1 + clockfree_audit_measurement_v1 + 8 Tier 2 passes) with zero regressions while unlocking two new required Tier 2 passes:
  • mode_hold (+6 view reqs): FAIL (hq=0.8411, 5/8 modes) $\to$ PASS (passing_suffix=14, hq=0.9993, 8/8 modes, cover=1.0).
  • five_word_joint_hold (+1 view req): FAIL (24/25 checks) $\to$ PASS (25/25 confirmed scheduled checks, mass_tv=0.0209, minimum_reconstruction_token_probability=0.9981).
• Imbalanced & Non-Stationary Metric Improvements:
  • vector_unequal_mass (held): preserves 2.91% mass on the 2% rare mode (min_mass_ratio=0.752) while lifting component_min_eigen_ratio from 0.0595 $\to$ 0.1605 >= 0.15.
  • gaussian1d_stability: improves final mean_error_sigma from 0.2869 $\to$ 0.0808 (still misses the intermediate abrupt-shift deadline).
• DCGAN Scaling Throughput (dcgan-performance-cloud-66339a696dff, NVIDIA RTX A6000, 54/54 uncontended cells):
  • 64², B64, W64: 39.93 ms (1,600.20 img/s) vs BISPF 89.38 ms (2.2× faster).
  • 256², B16, W64: 157.23 ms (101.73 img/s, 1,522 MiB) vs BISPF 342.19 ms (2,119 MiB, 2.2× faster / 28% less VRAM).

Verification

• python scripts/check_lint.py — 0 new findings
• python ops/refresh_forge_readouts.py --check --reports-only — CURRENT (stale stages: none)
• uv build — wheel 401,680 bytes (< 600,000 limit)
• pytest — 191 passed across test_empf.py, test_refresh_forge_readouts.py, test_forge_comparison_policy.py, test_forge_trainer_families.py, test_core_inventory_legacy_documentation.py, and test_recipe_defaults.py