WUT_Computer_Science/TODO-file-length-250.md

TODO: the 27 data .txt files the 250-line cap cannot classify

REMOVE ME AFTER FINISH

The 250-line cap applies to prose as well as code, and the shared gate treats .txt as prose unless its mean line is under 25 characters (a wordlist). These 27 files are data with long lines, so the heuristic reads them as prose and demands they be split. Splitting any of them corrupts it: in every one, a line is a record.

Nothing here blocks the refactor. It blocks exactly one step — installing the file-length pre-commit hook and workflow — so it must be settled before that lands, and not before.

Run bash ~/src/utils/scripts/check_file_length.sh --all to see the current list.

Why the existing mechanism does not cover them

VENDORED_SUBPATHS in ~/src/utils/file_length/_tables.py matches directory prefixes (/{repo}/{sub}/). That is enough for the 7 entries already in gates/vendored.txt, and it is not enough here:

• Programming/NLP/test_goldStandard/images/ holds the official SemEval corpus and kuhy's own GPT outputs, side by side.
• Programming/WDWR/projekt/R/ holds scenarios.txt (data) and model.R (authored, 464 lines, must stay gated and covered).

A directory entry would exempt authored work in both cases. Per-file # noqa style suppression is banned outright. So this needs a decision, not a workaround.

The four categories

1. Official SemEval package — 12 files, genuinely third-party

The STSint.testinput.{headlines,images,answers-students}.{sent1,sent2}[.chunk].txt files. Verified against the package's own manifest in Programming/NLP/test_goldStandard/00-README.txt, which enumerates exactly these, plus evalF1.pl and wellformed.pl, as the contents of the "STS 2016 Task 2 Interpretable Semantic Textual Similarity TEST DATASET" (authors: Eneko Agirre, Montse Maritxalar et al.).

Unambiguously vendored. The only question is how to express it.

2. kuhy's GPT outputs living inside the corpus directory — 6 files

Programming/NLP/test_goldStandard/headlines/headlines-chunks-gpt-{one,two}.txt
Programming/NLP/test_goldStandard/images/images-chunks-gpt-{one,two}.txt
Programming/NLP/test_goldStandard/student/students-chunks-gpt-{one,two}.txt

Not in the package manifest — these are output from gpt_chunks.py, written next to the corpus they were derived from. Regenerable in principle, but only by re-running non-deterministic paid GPT calls, so untracking them destroys the experimental record rather than saving space.

3. NLP pipeline data — 4 files

Programming/NLP/alignments_unformatted_headlines.txt   375 lines
Programming/NLP/alignments_unformatted_student.txt     344
Programming/NLP/output.txt                             348
Programming/NLP/reformated.txt                         348

Intermediate stage data. alignments_unformatted_headlines.txt is read as input by create_alignments.py:6 and format_alignments.py:7, so relocating these costs exactly two path constants — the cheapest category to fix.

4. WDWR solver data and output — 5 files

Programming/WDWR/projekt/old_code/data10000.txt        10000 lines  (numeric matrix, solver input)
Programming/WDWR/projekt/R/scenarios.txt                 999        (generated scenario vectors)
Programming/WDWR/projekt/wdwr/solutions1.txt            3038        (CPLEX MILP solver log)
Programming/WDWR/projekt/wdwr/KRZYSZTOF_.../scenarios.txt  999      (submission copy)
Programming/WDWR/projekt/wdwr/KRZYSZTOF_.../solutions1.txt 3038     (submission copy)

Two of these live under KRZYSZTOF_RUDNICKI_307585_WDWR_PROJEKT/, the submitted deliverable. Decision 6 of the approved plan says leave semantic duplicates alone, so relocation is not available for those two — moving a file inside a submission alters what was submitted.

Options

A. Extend the shared manifest to support file paths. Teach ~/src/utils/file_length/exemptions.py:is_vendored to read gates/vendored.txt at the repo root, supporting exact files as well as directories. This is the design originally chosen (one root manifest consumed by every gate) and is the only option that handles all four categories. Cost: file_length currently has zero tests and no coverage job in .github/workflows/shared-gates-tests.yml. Adding code to a gate 16 repos depend on means authoring its test suite first.

B. Relocate categories 2–4 into data/ subdirectories and add data to the manifest per course dir. Cheap for categories 2 and 3. Cannot touch the two WDWR submission-copy files. Does not address category 1.

C. Accept a documented residual — install the file-length gate with these 27 paths listed here, and treat the list as the record. Honest, but it is an allowlist by another name, which the fleet rules reject.

Recommendation: A, taken as its own ~/src/utils change with its own tests and CI job, sequenced before Milestone 2 step 4. B as a fallback for categories 2–3 if A is judged too large, leaving categories 1 and 4 to a narrower version of A.