Document GPU scheduling and regime-aware search

[magiccodingman]

Summary

• document measured shared-multi-GPU versus independent-per-GPU benchmark scheduling, candidate-aware NGL planning, topology caching, and the Qwen3.8 27B throughput result
• document controlled high-, mid-, and opt-in low-fidelity context strata, context rank pairs, effective-context rule scoping, and historical tensor vocabulary without historical-winner replay
• document source-aware cross-run identity, the distinction between a complete Pareto evidence archive and a curated release menu, and the corrected Qwen3.8 frontier audit
• update the prediction, existing-quant learning, nonlinear-survivor, README, and wiki-index pages so the new behavior is integrated rather than isolated

Research captured

• dynamic-v2 and dynamic-v3 sources are revision-pinned and treated as tensor vocabularies, not competing historical frontiers
• Q4 behavior is not assumed to transfer into IQ3 surroundings; beneficial, harmful, and suppression-only evidence is scoped to measured effective context
• the Qwen3.8 campaign produced three strict replacements over the published frontier and a 41-artifact nondominated source-aware union
• the measured two-RTX-3090 topology used shared GPUs for large/singleton work and independent workers below the measured crossover, with about 44% higher aggregate throughput on the representative Q5 pair

Validation

• all local Markdown file links resolve
• all v2 Markdown code fences are balanced
• git diff --check passes
• implementation terminology and configuration keys were cross-checked against Pipeline draft PRs #29 and #30

This remains a draft until the linked implementation PRs and wording receive final review.