Spec B: shard and verify an official ChaosNLI split

[AMBRA7592]

Summary

• add deterministic contiguous shard and offset/limit selection to the ChaosNLI adapter
• emit deterministic per-shard manifests and aggregate them with an exact ordered-UID coverage proof that rejects missing indices, gaps, overlaps, duplicates, and reordering
• scale cohort lookup and leave-one-out reliability without changing observable behavior
• add a Tier-2 harness and reproducibility evidence for the official ChaosNLI v1.0 SNLI split, while committing no licensed rows or converted outputs

Commit structure

1. 8f40405 — sharding, manifests, synthetic fixture/tests, and report harness
2. b6b9914 — independently revertible disagreement.py performance refactor plus seeded naive-vs-optimized equivalence test
3. cf3909a — evidence only: official-run manifest/report and reproduction instructions

The third commit is intentionally evidence-only: the committed manifest names the already-existing exact code commit used for the run (b6b9914). This preserves truthful provenance while keeping the two mandated code changes separate.

Tier 1 — CI-pinned logic

• 7 synthetic records over 3 uneven shards yield sizes 2/2/3 and exact ordered coverage
• offset/limit output is byte-identical to the corresponding shard ranges
• duplicate and gap/order corruptions are rejected after internally consistent manifest mutation
• broken converted distributions are rejected
• manifests reproduce byte-for-byte
• optimized reliability/cohort lookup equals the naive former implementation across forced ties, singleton cells, overlapping cohorts, and 250 seeded randomized configurations

Tier 2 — external empirical evidence

• official source: chaosNLI_snli.jsonl, ChaosNLI v1.0, CC BY-NC 4.0
• source SHA-256: 99f9015ddda7d85f66a087452bc30d53974314fe27e7d589e2f41ad44bd509c1
• 1,514 records / 151,400 anonymous votes / 16 shards
• distributions: 1,514/1,514 exact
• soft labels: 1,514/1,514 exact
• entropies: 1,514/1,514 at documented 3-decimal output precision
• schema validation and operational pipeline: 16/16 shards
• second complete run reproduced manifest.json byte-for-byte (1406173c…)
• reliability, error attribution, variation-vs-error, and cohort-fork prevalence are explicitly NON-APPLICABLE and undefined, not zero

Regression and scope gates

• default triage.json, soft_labels.jsonl, soft_labels.csv, and governance.jsonl are byte-identical to verified main
• resolution records are identical after removing runtime timestamps; replay hashes match exactly
• genuinely jsonschema-free venv: 31 tests, 29 pass + 2 intended skips
• jsonschema-enabled: 31/31 pass
• exact tracked scope: .gitignore, adapter code/docs/fixture, disagreement.py, test_claims.py, and reports/chaosnli/
• no official source row, UID, converted labels file, or virtual-voter record is committed

Audit note

This PR is intentionally left unmerged for independent release-gate review.