feat: preserve pinned RTPTorrent class execution evidence

[jacksonpradolima]

Problem and resulting behavior

RTPTorrent’s raw files contain job-level class executions, repeated class labels, reported seconds and partial method counts. Published baseline tables are filtered slices; treating them as complete CI history or inventing missing completion times can distort zero-failure coverage and leak outcomes.

Add a bounded, version-pinned source-assessment slice for #261. The typed dataset preserves project/job/class/position execution identity, provenance, exact reported seconds, explicit observed/unavailable duration status, partial method counts and duplicate-exclusion counts. It always exposes the unresolved temporal-admission blocker. The issue remains open.

Architecture and contract

• Reuse existing licensed MaterializationSpec, immutable external SourceCache, byte verification, ArtifactWriter, manifest, native runner and learning checkpoints.
• Add RTPTorrentSpec / RTPTorrentDataset schema v1 with recipe rtptorrent-class-results-v1, source format 1.1 and the exact eight-column raw layout. Filtered baseline layouts, unknown versions and implicit migration reject.
• Preserve repeated classes at distinct source positions; identical duplicate rows have explicit exclusion counts and conflicting same-key rows reject. Serialization order is canonical, never asserted chronological order. Source execution identity is distinct from cross-job logical test identity.
• Recheck the bytes actually parsed after cache preflight. Publication revalidates actual source correspondence, output containment and cache/output separation; concurrent retries converge and interrupted writes produce no completed artifact.
• Add one manifest artifact family. Core models/readers have no CLI, HTTP or MCP dependency. Existing result formats and checkpoint behavior retain their meaning.

Zero reported seconds remain observed zero. Empty duration is null/unavailable. Invalid negative/nonfinite/overprecision durations reject, including compact exponent amplification corrected in #288 / 4.4.1. Failure/error counts remain distinct from unique faults; skipped/unknown rows never become passes. Reported class elapsed time is not runner occupancy or parallel CI latency.

Use and reproducibility

from coleman.corpus import SourceCache
from coleman.corpus.rtptorrent import convert_rtptorrent_source
from coleman.corpus.rtptorrent_contracts import RTPTorrentSpec

# receipt comes from the existing materializer with exact bytes and reviewed data terms.
cache = SourceCache("/absolute/path/outside-git/source-cache")
spec = RTPTorrentSpec(schema_version=1, materialization=receipt,
                     member="classes.csv", project="owned/ci", history="constructed")
dataset = convert_rtptorrent_source(spec, cache)
# dataset.require_temporal_evaluation() raises: event/outcome clocks and build mapping unavailable.

docs/rtptorrent.md supplies the complete owned materialization/conversion example, primary-source assessment, exact source/member identities, compatibility rules, resource bounds and Mermaid diagram. examples/rtptorrent/ci.py runs one owned constructed isolated serial job through the real Random/RNFail native pipeline, checks independent exact metric references and reopens its checkpoint. Installed-wheel testing copies/runs it twice outside the checkout and executes the complete documentation example.

Manual source check (network-free conversion after acquisition): the exact HikariCP raw member produced 46,995 executions / 1,662 jobs, retaining 23,302 repeated-class executions, no duplicate exclusions and 1,530 jobs without reported class failures/errors. Its filtered untreated baseline contains 4,611 rows / 125 jobs and no such green jobs. These are descriptive source counts, not independent verification of whole-job success or policy effectiveness. Member digests are not archive digests. No third-party source rows are committed.

Validation and local review

• Full suite: 2,035 passed, 13 existing telemetry warnings; 95.31% combined line/branch coverage. Command: pytest -n4 --cov --cov-branch with real Syft available.
• 63 direct adapter unit cases, plus a noneditable base-wheel installation outside the checkout, import-boundary checks and real native TCPCI/checkpoint replay. The full suite and all quality checks were repeated after the consolidated Sonar corrections.
• Round trips and versioned owned golden fixture; legacy/future/schema/unit rejection; zero/null/unavailable/nonfinite/negative handling; deterministic identities/digests; row/decimal/CRLF metamorphic checks; duplicate conservation; concurrent/idempotent/interrupted writes; mutated-cache byte checks; CSV framing/resource limits.
• make lint format-check typecheck interrogate cost-structural docs build, pylint -E coleman and git diff --check: passed.
• Alibaba OpenCodeReview 1.12.13 local delegation only: workspace/full main-to-HEAD previews and resolved rules. 11/11 files reviewed, five tool-selected files plus six excluded files manually reviewed; golden values were manually inspected too. No external model/review endpoint was used; no confirmed critical/high/medium issue remains.

Compatibility, migration and limits

No existing artifact is reinterpreted or automatically migrated. Source JSON checks model invariants; validate_rtptorrent_dataset checks correspondence against actual pinned cache bytes. Recipe-only converted rows remain local. Input bounds are 16 MiB per selected member, 50,000 physical data rows, 16 KiB per physical line and 1,024 characters per field and 64 characters per duration literal; these are not an OS memory sandbox.

Temporal admission remains blocked: raw tables lack event/completion clocks and complete build/job mapping, and repeated class labels lack module/configuration disambiguation. The original paper’s metadata snapshot differs from the accessible dump; join equivalence/coverage remain unverified. Full RTPTorrent archive acquisition is incomplete; its ZIP64 layout / 2,883 entries exceed the existing v1 materializer contract, which is unchanged. The original paper also reports negative runner-duration defects; this reader rejects those rows without silent repair.

The registry/diversity matrix and its seven uncovered real-evidence regimes remain unchanged. The owned native slice establishes functional compatibility, not an observed CI timeline, generalization, policy superiority or elapsed-time savings. Subsequent #261 work covers LRTS, controlled fault corpora, configurable/synthetic stress, comparable baselines/statistics and protected confirmation/reporting.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-16T02:32:46.724662Z	13f5d05	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[jacksonpradolima]

@codex review

Consolidated corrections in b5092e2 address the seven initial Sonar findings: simplified numeric regexes retain explicit re.ASCII (Unicode digits remain invalid), and the owned CI example delegates output containment/cache separation to one helper. Seventeen new lexical-boundary cases verify accepted decimal spellings and malformed/Unicode rejection. Final full suite: 2,034 passed, 95.30% combined coverage; installed native replay/checkpoint and all static/Pylint/docs/build checks pass. The real pinned 46,995-row HikariCP member was reconverted successfully, retaining repeated classes and the temporal blocker.

Repeated local OpenCodeReview 1.12.13 workspace and full main-to-HEAD delegation/rules review covers all 11 files: five selected, six manual exclusions, including full fixture values. No confirmed critical/high/medium issue remains locally. Please review the current head.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Hooray!

Reviewed commit: b5092e2258

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[jacksonpradolima]

@codex review

13f5d05 resolves the additional S8786 finding by matching mantissa and exponent separately, removing ambiguous backtracking while preserving the ASCII-only grammar. A regression checks that 64-character zero is accepted and 65-character zero rejects; the documented resource limit is explicit. The original seven Sonar findings were already corrected in b5092e2.

Final full suite after this correction: 2,035 passed, 95.31% combined coverage. All 63 adapter unit cases, the installed-wheel/native-checkpoint workflow, static typing, lint/format, Pylint, docstrings, structural checks, strict docs and builds pass. Reconverted the real pinned 46,995-row source successfully with admission still blocked. Repeated OpenCodeReview 1.12.13 local delegation workspace/full-range review and manual exclusions cover all 11 changed files; no confirmed critical/high/medium finding remains locally. Please review the current head.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
1.2% Duplication on New Code

See analysis details on SonarQube Cloud

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Keep it up!

Reviewed commit: 13f5d05a7a

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".