Product-Readiness and Architectural-Boundary Closeout of the Spiral-Domain H-Pipeline: A Pre-Registered Modal-Cloud Campaign Establishing 23-Hour Sustained Streaming, Empirical Differential-Privacy Defensibility, a Causally-Proven Class-Imbalance Artifact, an Input-Dependent Compression Tax With Two Falsified Operating-Point Levers, and Five Hybrid-Architecture Deployment Stages

This paper reports Phase XXXI of the spiral-domain H-pipeline campaign — the product-readiness and architectural-boundary closeout — executed on Modal cloud A100. It comprises a core operational-validation sweep and a gap-closure extension, with every sub-campaign frozen in a pre-registration scope document before any data examination, pre-committed Outcome A / B / C decision rules, and verbatim honest-negative reporting. The cumulative campaign across Phases I–XXXI now exceeds 2,000 pre-registered studies (study identifiers through 2113). Product-readiness anchors. · A 23-hour sustained streaming soak: 185,803,300 ECG records at 2,244 records per second, with 99th-percentile latency improving from 0.522 ms (Hour 1) to 0.513 ms (Hour 23) within a 0.512–0.558 ms band, −0.84 MB memory drift, zero errors, and 28 cumulative clean hours across two independent days. · Empirical differential privacy: a capable ensemble membership-inference adversary scores AUC = 0.483 — below the 0.50 chance line — at σ = 1.0 codebook-centroid noise on an n = 5,000 cohort. · Concurrency: 99th-percentile latency stays flat (0.479–0.499 ms) from 1 to 500 concurrent streams on a single A100. · Multi-modality: ECG and kinematic streams are ingested concurrently at sub-millisecond per-modality latency. · Privacy at scale: membership-inference AUC falls from 0.62 (n = 1,500) to 0.52 (n = 5,000) — privacy strengthens as the cohort dilutes the signal. · Cohort scaling: a deeper-LoRA configuration recovers the multi-class headline at n = 5,000 (macro-AUC 0.6547, statistically tied with the Chronos comparator 0.6613, Δ = 0.007 within seed variance), with an honestly-disclosed joint encoder/dataset ceiling at n = 10,000 affecting all encoders. Architectural-boundary findings (the scientific spine). · Class-imbalance artifact, causally proven: the extreme-scale macro-AUC decline (to ≈0.540 at n ≈ 14,000) is a PTB-XL class-imbalance artifact — re-balancing the cohort recovers macro-AUC to 0.621 (causal), and a pretrained foundation-model comparator declines identically to 0.561 (encoder-agnostic). It is not an architectural ceiling. · Input-dependent compression tax: the one-token compression tax versus the best same-input uncompressed classifier is ≈0.02–0.05 macro-AUC on information-poor single-lead input but grows to 0.134 on information-rich 12-lead input (H-pipeline 0.700 versus a same-input RandomForest ceiling of 0.835). · Two operating-point levers FALSIFIED: two pre-registered attempts to recover the 12-lead tax by emitting more tokens — temporal segmentation (ΔAUC(S = 8 − S = 1) = −0.057) and per-facet "multi-twin" discretization (ΔAUC(G = 12 − G = 1) = −0.093) — both reduce accuracy, establishing that the binding constraint is the feature representation, not the token count. · Privacy matured to a threat-model-plus-defense: a real white-box within-session linkage threat (AUC = 0.920) is driven to chance by codebook-centroid noise (σ = 4.0 → 0.518, p ≈ 0); the noise scale is a tunable dial calibrated to the threat. Five hybrid-architecture deployment stages are introduced and empirically anchored: Stage M-A drift detection (codebook-histogram Jensen-Shannon divergence; out-of-source AUC = 1.000; real cross-dataset AUC = 1.000 with a tight bootstrap interval); Stage M-B edge-cloud triage with host-CPU/A100 streaming latency parity (0.385 ms ≈ 0.389 ms, ratio 0.989) — whose confidence-gated routing embellishment is disclosed as an honest scope boundary because the maximum-compression representation yields well-ranked but under-confident per-record probabilities; Stage M-C foundation-model-embedding-as-feature-extractor (end-to-end 5-class macro-AUC 0.639, within 0.018 of native); Stage M-D multi-task heads from one shared token stream; and Stage M-E production-load robustness under co-tenant GPU contention (99th-percentile 0.212 / 0.284 / 0.138 ms across no / realistic / aggressive contention, zero-percent throughput degradation). The unifying conclusion is that the H-pipeline is compression, energy, latency, privacy, and drift-detection infrastructure for multi-class clinical triage in bandwidth-constrained deployment — not a per-record binary classifier, and not a system whose accuracy is tunable by token count. Companion to Papers 12 (Parent H 64/080,096), 13 (Parents I/J/K), 14 v3 (Parent L 64/084,827), 15, 16, and 17. The operational and architectural-boundary findings are covered by the filed Parents H/I/J/K/L; the hybrid-architecture deployment patterns (Stages M-A through M-E) are covered by an additional provisional, Parent M (64/090,833), filed June 15, 2026. PYTHONHASHSEED = 0 deterministic; Phase XXXI per-study summary JSONs, harness library, pre-registration scope documents, and Modal runners are included in the archive. Companion deposits in this Zenodo Community (spiral-domain-encoder-campaign): · Paper 13 — 10.5281/zenodo.20595834 · Paper 14 v3 — 10.5281/zenodo.20596597 · Paper 15 — 10.5281/zenodo.20596931 · Paper 16 — 10.5281/zenodo.20602959 · Paper 17 — 10.5281/zenodo.20634763 · Paper 18 — 10.5281/zenodo.20709337(https://doi.org/10.5281/zenodo.20709337) ← this deposit Licensing inquiries: Randolph James Ferlic, M.D., randolphf@fieldstoneanalyticsllc.com.