Skip to content

COLEMAN 4.8 — Real CI history, reproducible test prioritization runs

Turn a pinned RTPTorrent archive into inspectable Parquet and checkpointed TCPCI experiments with the runner you already use. This release connects verified source acquisition, complete analytical history and a bounded native replay in one documented workflow.

What you can do

  • Prepare real RTPTorrent CI observations. The HikariCP preset preserves 46,995 class executions across 1,662 source jobs in analytical Parquet, with exact reported seconds, original execution identities and explicit missing values.
  • Run and resume Random/UCB experiments. An explicit eight-job preset projects 18 compatible executions into the existing native API. Two policies produce 16 result rows across eight modeled rounds; retrying resumes committed checkpoints without duplicating results.
  • Inspect every exclusion. The other 46,977 executions remain analytical records. Source, projection and duplicate counts must reconcile, so a smaller native slice never erases the original history.
  • Verify what you downloaded. A separate bounded ZIP64 profile verifies the full approximately 5 GB source, selected-member membership and decoded hashes. Resumable operator downloads and an external cache keep large data out of the repository and wheel.
  • Reproduce artifacts deliberately. Versioned recipes bind source terms, hashes, assumptions, actual producer code and Parquet writer version. Immutable tables precede the completion receipt; interruption, concurrent retries and unknown versions have explicit behavior.

The release also corrects Linux memory accounting for application workers: a large parent process can no longer make a small worker appear over its RSS limit. The guard measures the worker executable's kernel high-water mark, retains existing admission limits and fails closed when measurement is unavailable. Real Linux regressions cover both native/coverage execution and actual low-cap refusal.

Try it

Use Python 3.14+ and install the verified 4.8 wheel from the release assets. After downloading the pinned archive to an external workspace:

python -m coleman.transports.rtptorrent_etl \
  --archive /external/rtp-torrent-v11.zip \
  --cache /external/coleman-cache --output /external/coleman-data --run

# Once cached, verify and resume offline.
python -m coleman.transports.rtptorrent_etl \
  --cache /external/coleman-cache --output /external/coleman-data --run

See the RTPTorrent Parquet and replay guide for installation, source download, offline replay and the typed API. A small repository-authored example exercises the same complete path without downloading foreign data. The existing LRTS workflow remains available.

Evidence you can inspect

The installed-package workflow runs outside checkout against the real source cache, validates Parquet, executes the actual native runner twice and checks committed checkpoint recovery. Local regression coverage includes serialization/goldens, unknown versions, incompatible units, zero/null/unknown/not-run semantics, source-order metamorphism, duplicate conservation, concurrent exports, interrupted writes, forged provenance and import boundaries. Complete local review and CI evidence are linked from the implementation PRs.

Scope and compatibility

This delivery supports post-hoc source-job serial simulation. Numeric job labels are not verified chronology; reported class seconds are not parallel CI wall-clock savings; repeated seeds are not independent observations. Protected temporal admission remains blocked by missing source clocks and logical build/job mapping. The broader diversified corpus and confirmatory comparisons in

261 remain ongoing.

Existing ZIP recipes, raw readers, native Parquet, LRTS exports and checkpoints retain their semantics. No historical artifact is silently migrated. New receipts require explicit schema/producer/writer compatibility. Source data terms remain separate from enclosed subject-code rights, and the workflow does not execute archived subject repositories.

Implementation and review: ZIP64 source integrity (#314), real Parquet and native replay (#315). The complete local suite passed 2,542 tests with 95.31% coverage, with the Linux-only regression separately exercised in Linux and an installed wheel. Full changes.