Skip to content

COLEMAN 4.9 — Larger CI histories, explicit test stages

Move beyond a single real-data demonstration while keeping every stage and observation inspectable. COLEMAN now prepares four-build ActiveMQ and Log4j histories as analytical Parquet and runs their declared native stage views through the same TCPCI API, policies and resumable checkpoints.

What you can do

  • Use larger real method histories. Frozen source recipes preserve 18,901 ActiveMQ observations and 30,066 Log4j observations with original PR/build, suite, class, method and execution identities.
  • Keep environments distinct. Separate Log4j Ubuntu and Windows native views bind exact raw context arrays and build-local node IDs to processed stages. All stages remain in analytical files; selecting one view never erases the rest.
  • Audit every native exclusion. Outside-stage, not-run, unknown-outcome and unavailable-duration records retain their original evidence and reconcile with the complete source. Reported zero seconds remains distinct from missing data.
  • Run, inspect and resume. Each four-build view runs Random and UCB with the existing native RNFail feedback and checkpoint protocol. Immutable source and completion receipts make identical retries reproducible.
  • Recognize missing evidence. Typed temporal blockers distinguish resolved within-report stage correspondence from unavailable executor/outcome/feature clocks, expected test universe, verified continuity and concurrency topology.

Use the complete stage-view guide for exact source downloads, bounded local acquisition, Parquet schemas, typed API, native runs, recovery and attribution. Large third-party archives stay outside Git and the wheel. Recipes independently verify whole-archive, nested-member and decoded-report bytes; no archived subject repository is executed.

Compatibility and evidence

The new versioned recipe family has its own producer and artifact namespace. Existing single-stage LRTS recipes/receipts, raw readers, RTPTorrent workflows, RunSpec identities and checkpoints retain their semantics. Old artifacts are not silently migrated. Unknown schemas, producers or Parquet writers reject with an explicit compatibility requirement.

Stage views are post-hoc serial simulations. Reported scheduled labels are not verified outcome chronology; method-duration sums are not physical parallel CI savings. Both Log4j views share the same source project/build groups and do not become independent projects or faults. The profiles retain blocked_source_evidence and insufficient_evidence; the diversified corpus, protected comparisons and confirmatory report in #261 remain ongoing.

Validation covers the installed base wheel outside checkout, actual real-source Parquet/native/checkpoint workflows, versioned goldens and old artifacts, round-trips, units/missingness, non-finite rejection, collision guards, forged provenance, deterministic identities, duplicate conservation, interrupted writes, concurrent retries and domain/transport import boundaries. This is evidence of a reproducible pipeline, not an algorithm effectiveness or production-savings claim.