RTPTorrent: source assessment and bounded class-result conversion¶

The RTPTorrent adapter in COLEMAN 4.5 converts raw class-result CSV members, using the existing licensed materialization/cache contract. The verified 4.9 wheel retains this reader and includes the complete real Parquet/replay workflow. RTPTorrent remains blocked from protected temporal evaluation.

An execution is identified by project, job ID, class label and source position. Repeated class labels remain separate executions. Job IDs are neither logical build IDs nor timestamps. No commit date, synthetic clock or sum of durations is substituted for outcome availability.

Research questions and source decisions¶

The source audit asks which exact bytes and terms can be pinned, what an observation means, whether feedback was available before a decision, and which evidence a baseline table actually retains. It uses the original paper, artifact README and actual bounded member inspection. This is a focused source-admission assessment, not a completed systematic literature review or an effectiveness study.

Primary sources checked on 2026-10-15/16 UTC:

• RTPTorrent 1.1, DOI 10.5281/zenodo.4046180, and its archive README. The provider advertises rtp-torrent-v11.zip, 5,016,963,482 bytes, MD5 1f7fa822b0cf155bd007a94d1a24a336, and CC BY 4.0 metadata.
• The authors' MSR 2020 paper describes the historical Java/Travis corpus and its limitations. The README defines class-level duration in seconds and separate method failure/error/skip counts, and warns against using concurrent sibling jobs as prior feedback.
• TravisTorrent's archive notice directs users to Figshare record 19314170 and states that raw-log access ended in 2022. The checked record is version 1; final-2017-01-25.csv.gz has 264,607,174 bytes and provider MD5 5d359345fecf0bf47be6edb10b5aae58.
• TravisTorrent field documentation describes its trigger timestamp. The inspected 66-column CSV header contains no explicit completion field. Matching coverage and full dump integrity remain unverified. Travis API documentation distinguishes start/completion and warns that restarted build duration need not equal their difference. These fields cannot silently create outcome times.

The original paper names travistorrent_8_2_2017.csv as its metadata join table. The inspected Figshare dump is final-2017-01-25.csv.gz. Their equivalence and join coverage have not been verified; the accessible snapshot cannot silently substitute for the one cited in the paper.

The paper also reports 17 negative durations caused by test-runner defects and notes that clock resolution produces reported zeros. This adapter rejects negative duration rows rather than repairing them or treating them as unavailable. A valid reported zero remains zero; it does not prove that execution consumed no physical time. Changing logging, platforms, components and hardware further limit cross-job identity and timing comparisons.

The RTPTorrent ZIP64 end record advertises 2,883 entries and 395,312 central-directory bytes. This exceeds the existing v1 materializer's entry ceiling and ZIP64 policy. That contract remains unchanged. The implementation delivered in 4.8 through #301 adds a separate bounded ZIP64 selection profile, without changing legacy recipes. The profile is available in the published 4.9 wheel. A real selection recipe still requires the complete archive SHA-256; relabeling a partial archive as an already validated ZIP would bypass the existing contract.

Observed source discrepancy¶

The following single-project descriptive check inspected the HikariCP raw member and its published untreated baseline member. It did not pool projects or estimate independent observations:

The raw member has 23,693 distinct job/class pairs but 46,995 distinct job/class/position keys. Collapsing by class would lose executions. The published baseline is a filtered slice; it cannot stand in for the complete raw history or establish a zero-failure regime. Counts of source rows without reported failures are not independent verification that the entire CI job passed.

The independently computed raw-member SHA-256 is 0ebc3dd28d09227f6aa56698b76c7a217d034aa83318ffc5a75ff8d2d0632687 for 2,825,388 decoded bytes. The baseline-member SHA-256 is ad162fbdd9786db59a30b7bc5625fab19cbac9d5b11e47ca6c1d79b46e041fa9 for 339,235 decoded bytes. These are member identities, not the archive SHA-256. Initial bounded HTTPS ranges, ZIP metadata and member CRC/size/deflate-EOF checks informed this research. On 2026-10-16 complete acquisition verified all 5,016,963,482 bytes and the provider MD5; complete source SHA-256 is bff52563d75aee2f4bab9693188cdefbfd887dc4bcf63395d49eaff1ecf7b6c1. The new bounded ZIP64 profile verified actual enclosing membership and repeated offline conversion of the raw HikariCP member. Its frozen metadata recipe is included in the published 4.9 wheel. The real Parquet ETL builds on that profile, and source integrity does not supply missing temporal evidence. No third-party source rows or archive are committed here.

To reproduce the descriptive counts, obtain the exact 1.1 archive outside Git, verify its complete bytes against the provider's size/MD5 and compute its SHA-256, then export the named members with a separately reviewed ZIP64-capable tool. Pin each decoded member before conversion. Compare raw row counts, distinct job IDs, distinct job/class and job/class/position keys, and group-level absence of positive failure/error counts. Do not treat a partial archive or range checksum as proof of the entire container. Complete archive acquisition/validation is still an outstanding reproducibility step, not a passing COLEMAN fetch recipe.

Admission decision: blocked. The raw layout supplies no event/completion clock or complete build/job mapping, and repeated class labels lack a disambiguating module/configuration dimension. A frozen completion/mapping source and explicit identity/outcome rules are needed before temporal baselines or confirmation. This is a documented compatibility limit, not a claim that the corpus is permanently unusable. The registry and all seven uncovered diversity regimes remain unchanged.

Contract and compatibility¶

flowchart TD
    S[Explicit source bytes and reviewed terms] --> M[Existing pinned materialization]
    M --> C[External immutable source cache]
    C --> R[Bounded raw class-result reader]
    R --> D[Versioned execution dataset and provenance]
    D --> B[Temporal evaluation blocked: clocks and mapping missing]
    D --> A[Source-correspondence check and immutable local artifact]
    F[Owned constructed fixture only] --> I[Isolated serial native projection]
    I --> N[Existing runner and learning checkpoint]
    N --> K[Independent known-answer check]

RTPTorrentSpec and RTPTorrentDataset require integer schema_version=1 and the rtptorrent-class-results-v1 recipe. Missing, legacy, extra-field and future contracts reject; there is no implicit migration from native Parquet, baseline CSV, source cards or materialization receipts. Existing artifacts retain their meaning. validate_rtptorrent_dataset recomputes conversion from verified cache bytes; reading a JSON artifact alone checks invariants, not source correspondence.

The accepted header is exactly:

travisJobId,testName,index,duration,count,failures,errors,skipped

Durations use s with meaning rtptorrent-reported-class-elapsed-v1. They are not observed runner work or parallel elapsed CI savings. Counts are nonnegative integers; known failing/error/skipped counts cannot exceed the total. No unique fault, calibrated probability, flakiness or independence is inferred. The frozen spec retains duration units even for null values. Non-null duration quantities require duration_status=observed; this source layout supplies no not-applicable duration encoding, so that status rejects rather than being invented.

Limits are 16 MiB per selected member, at most 50,000 physical data rows, 16 KiB per physical line and 1,024 characters per field. Duration literals have a separate 64-character ceiling before Decimal construction. Lower row/line limits may be frozen in the spec. UTF-8 comma-separated single-line records are required; multiline quoted labels, malformed records and filtered baseline headers reject. These are bounded input dimensions, not an operating-system memory sandbox.

Complete owned example¶

Install the verified 4.5 release wheel in an activated Python 3.14+ environment. Run this example outside the checkout. The license below applies only to this newly authored fixture; never reuse it for downloaded data.

import hashlib
import io
from tempfile import TemporaryDirectory
from pathlib import Path

from coleman.artifacts import ArtifactWriter
from coleman.corpus import MaterializationSpec, SourceCache, load_registry, materialize_source
from coleman.corpus.rtptorrent import convert_rtptorrent_source, publish_rtptorrent_dataset
from coleman.corpus.rtptorrent_contracts import RTPTorrentSpec

payload = (
    b"travisJobId,testName,index,duration,count,failures,errors,skipped\n"
    b"101,owned.Test,0,0,1,0,0,0\n"
    b"101,owned.Test,1,2,1,1,0,0\n"
)
digest = "sha256:" + hashlib.sha256(payload).hexdigest()
package = MaterializationSpec.model_validate({
    "schema_version": 1,
    "package_id": "owned-rtptorrent-format",
    "archive": "raw",
    "source": {
        "uri": "urn:coleman:owned:rtptorrent-format",
        "version": "1.1",
        "citation": "Repository-authored format-compatible CI fixture",
        "sha256": digest, "size_bytes": len(payload), "availability": "not_verified",
    },
    "license": load_registry().card("ci-serial-boundary-v1").manifest.license,
    "members": [{"name": "classes.csv", "sha256": digest, "size_bytes": len(payload)}],
})
with TemporaryDirectory() as temporary:
    root = Path(temporary)
    cache = SourceCache(root / "external-cache")
    receipt = materialize_source(package, cache, lambda: io.BytesIO(payload))
    spec = RTPTorrentSpec(schema_version=1, materialization=receipt,
                         member="classes.csv", project="owned/ci", history="constructed")
    dataset = convert_rtptorrent_source(spec, cache)
    assert len(dataset.executions) == 2 and dataset.repeated_class_executions == 1
    artifact = publish_rtptorrent_dataset(dataset, cache, ArtifactWriter(root / "exports"))
    assert artifact.exists()
    try:
        dataset.require_temporal_evaluation()
    except ValueError as error:
        print(error)  # Expected: no source clock or logical-build mapping.
    else:
        raise AssertionError("A clockless source entered temporal evaluation")

For native execution, copy the complete owned CI example outside the checkout and run:

python ci.py owned-ci /absolute/path/outside-git/rtptorrent-cache
python ci.py owned-ci /absolute/path/outside-git/rtptorrent-cache

It materializes and converts an owned format-compatible fixture, preserves the reported green/repeated-class and unknown jobs, and projects only one complete, unique-class constructed isolated serial job to native Parquet. Random/RNFail run through the existing runner. Independent exact references check its native metrics, and reopening the learning checkpoint preserves one result row. This example uses no cross-job learning and claims no observed CI timeline or savings.

Converted JSON contains source rows: keep recipe-only data local, outside public commits and mirrors. The publisher validates exact cached source correspondence, cache/output separation and output containment before immutable publication. Identical concurrent retries converge; interrupted writes produce no completed JSON. Process-kill temporary-file and filesystem durability limits are inherited from the existing artifact writer. No separate experiment framework, learning checkpoint, CLI dependency, MCP dependency or automatic download is introduced.