ProgramBench csview: false not_run records and missing results.xml

[andy6011960]

Environment

• ProgramBench: 1.2.4
• Repository commit: 963063c9271cc40fa179977356782ea4582e0b0c
• Instance: wfxr__csview.8ac4de0
• Cleanroom image: programbench/wfxr_1776_csview.8ac4de0:task_cleanroom_v6
• Image digest: sha256:e885e345ca055180d8d5579df46cc62eedf818f00aeab5712f4a6f456eb92fbd
• Host: Ubuntu 24.04.4 LTS, x86_64, native Docker
• Inference container: --network none

Reproduction

The same submission was evaluated twice:

uv run programbench eval ~/SINTEZ/PROGRAMBENCH_EVALS/csview_v0_5

uv run programbench eval \
  ~/SINTEZ/PROGRAMBENCH_EVALS/csview_v0_5_serial \
  --docker-cpus 1 \
  --branch-retries 3

Both runs produced the same raw result counts:

• passed: 344
• failed: 1
• skipped: 1
• not_run: 155
• displayed score: 68

The generated executable SHA-256 was identical in both evaluations:

a86228c9fd2d3c8d8dc25fa409418b8e48115b9dbf0ed5072b8797afa1e3d888

Issue 1: test-name prefix mismatch creates 153 false not_run records

For branch d31025219260, the evaluator reports both:

• 153/153 expected tests missing from JUnit XML
• 153 test(s) in JUnit XML not in tests.json

The records form 153 exact pairs after removing only the leading eval. prefix. For example:

• metadata: tests.test_edge_cases.test_all_cells_empty
• JUnit: eval.tests.test_edge_cases.test_all_cells_empty

All 153 JUnit-side records passed, while the metadata-side names were counted as not_run.

Suggested fix: normalize an optional leading eval. prefix before matching JUnit names to tests.json, or generate both sources with the same package root.

Issue 2: branch efa8c407dbe3 never produces results.xml

The branch fails with:

results_read_failed: Could not find the file /workspace/eval/results.xml

This occurred in the default run and again in the serial run. With --branch-retries 3, the initial attempt plus all three retries failed identically. Two expected tests remain not_run because no JUnit result file is available.

Suggested fix: preserve and expose the branch container's test-runner stdout/stderr when results.xml is missing, and ensure the runner emits a minimal JUnit file even on setup or collection failure.

Issue 3: --help golden differs from the cleanroom reference behavior

The only completed-test failure is:

eval.tests.test_cli_basics.test_help_exact

However, in the official cleanroom container, a byte-for-byte comparison between the preserved reference executable and the reconstructed executable produced no diff:

diff -u \
  <(/workspace/executable --help) \
  <(/candidate/csview_reconstruction_v0_5/executable --help)

The local cleanroom differential suite also passed 49/49, including the exact help behavior. This suggests that the branch golden may not correspond to the preserved reference executable in task_cleanroom_v6.

Suggested fix: regenerate help.txt directly from the preserved reference executable used for this instance, or document which reference/version the golden represents.

Why this matters

Among completed scored tests, the observed pass rate is 344 / 345 = 99.710145%. The displayed score of 68 is materially affected by evaluator/metadata artifacts: 153 prefix-mismatched names and two tests in a branch that never emits results.

This is not a request to rewrite an official score. It is a request to inspect and correct the evaluation artifacts so that the benchmark result reflects the tests that actually ran.

Public evidence attached

The attached public-safe evidence archive contains:

• both raw evaluation JSON files;
• a machine-readable audit summary;
• the deterministic audit script;
• the full bilingual technical report;
• SHA-256 checksums.

No reconstruction source code, executable, or submission archive is included in the public attachment. The reproducible candidate submission can be provided privately to maintainers if needed.

programbench_csview_public_issue_evidence_v1.zip