Add audited Crane scheduling, real workloads, experiments and an offline demo

[aHappend]

The old experiment paths used hand-authored chain proxies and could change total arithmetic when sub-batch size changed. Multi-parent traffic and hierarchy boundary fallbacks also made quantitative results hard to trust. This change adds an audited independent reproduction workflow with real graph metadata, recorded native core costs, raw experiments and an offline interactive demo.

Change

• Import 16 compiled SET network definitions with shapes, operation counts, residual/dynamic-weight edges and provenance. Provide recorded native SET Polar intra-layer mappings for four benchmark networks.
• Correct batch-work conservation, sample-index traffic identity, full multi-parent demand, block-boundary tensor volumes and DAG handling.
• Add normalized exact integer-product ScT optimization, an independently checked canonical vertex reduction, bounded solver calls and status/bound/gap records.
• Replace the default hierarchical path with parent-sub-batch macro composition, actual tile limits, explicit activation budgets and serial alternatives. Keep the old training MILP marked experimental and add a checked uniform-cohort training reference.
• Add isolated experiment matrices, seeded native SET runs, evidence verification/archive, figures and a self-contained responsive demo. Generated data are identified in Git attributes; cross-platform line endings preserve profile hashes.

Validation

• 46 tests passed locally, including independent small-instance enumeration, workload/gradient coverage, graph preservation, tensor intervals, nested resource/batch checks, profile accounting and rejection of tampered evidence.
• Ruff, dependency checks, citation schema, documentation links and GitHub YAML passed.
• 10 real-workload inference comparisons completed with all modeled checks passing.
• 9 training capacity cases produced 7 feasible outcomes and 2 explicitly expected policy-infeasible outcomes.
• 12 native SET runs completed: four networks, three fixed seeds, 20 SA rounds per layer, four internal trials. Original summaries, trees, parameters and seed patches are archived.
• The checked-in raw archive can regenerate the report, plots and demo without running native SET again.
• Headless Chromium verified desktop/mobile layout, filters, state playback, training outcomes and the SET tab without page errors.

Documentation and interpretation

The bilingual README, mathematical audit, architecture, reproduction map and experiment guide describe the final implementation and its boundaries. The result report retains VGG-19's slight regression relative to the same-model serial reference.

This is not a claim of reproducing all original paper figures or headline speedups. Python's outer traffic/placement model is not fully calibrated to SET; native SET totals are reported separately. Training evidence uses a conservative serial uniform-cohort reference and excludes the full per-layer checkpoint search and optimizer-state capacity. These limitations are explicit in both documentation and the demo.

Remote validation: all three CI jobs passed on the final head b29f0449b03e7ff0c72b532f9d6b229482c29ea0: https://github.com/aHappend/Crane/actions/runs/35624135131. Merge commits preserve the source revisions referenced by the experiment manifests.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.