feat: add communication phase regression scoring

[0z5a]

Draft update — 2026-09-05

Adds strict communication-evidence hard-negative gates: malformed or unbound evidence cannot produce a passing policy score. The later native-import stack passed 397 CPU tests and 138 schema-freshness checks; these are software checks, not independent GPU hardware confirmation. The baseline entrypoint repair is supplied by dependency #2; the old missing-entrypoint note below is historical.

AI assistance: OpenAI Codex. Remains Draft; implementation, CPU/Gloo checks and production GPU validation are separate acceptance gates. No new GPU results were generated during publication.

The original submission description and validation history follow.

---

Summary

• add framework-neutral schemas for communication trace, policy, and regression-result artifacts for arbitrary world_size >= 2
• validate communicator-local rank sequence and completeness, including overlapping communicator groups
• score GPU-realized phase offsets, pair completion/stretch, rank skew, consumer slack, and outstanding bytes/count without collapsing unrelated workload cells into a global score
• compare only exact system/path/load cells and map the new evidence into the existing SystemPathLoadCell model
• keep kernel-observed timing distinct from event-bracket timing

Why this is not duplicate work

I searched open PRs and issues for communication phase regression scoring and collective overlap regression gates. No existing change provides this framework-neutral artifact contract and strict same-cell comparison path.

Validation

• new CPU tests with logical 2-rank and 4-rank fixtures: 8 passed
• full repository suite: 349 passed, 1 deselected
• Ruff passed across src and tests
• all 136 schemas regenerate without drift
• one unrelated baseline-only failure remains in tests/test_historical_bulk_gates.py::test_training_campaign_uses_large_groups_after_safe_boundary: upstream main references the untracked file benchmarks/historical_prs/infraswe-inference-bulk-supervisor.sh; this branch does not touch that area
• this PR makes no hardware performance claim

AI assistance disclosure

This draft was developed with OpenAI Codex assistance. The human submitter must review and understand every changed line before marking it ready for review.