WUT_Computer_Science/CLAUDE.md

WUT_Computer_Science

Archive of university coursework from Warsaw University of Technology. 44 course directories, 12 languages, ~28k lines of authored code plus the submitted reports, figures and datasets. It is an archive: the code is finished, and the value is that it still builds, runs and is readable.

Repository Layout

Path	What it is
Programming/	25 course dirs with real code (Python, C/C++, Java, TS/Angular, C#, Unity, MATLAB-adjacent).
NotProgramming/	19 course dirs that are mostly LaTeX reports, PDFs and EDA schematics. Exceptions: ENUME (MATLAB), STUP/makieta (Angular), EDABA-LAB (SQL).
gates/	Machine-readable gate inputs. vendored.txt is the source of truth for "not my code".
scripts/	Repo-local gates. Shims that exec into ~/utils live here too.

Two structural facts that surprise people:

• Programming/PSD/zin3/third/ and Programming/psd_project/ are the same services under different names, and WDWR/project/ vs WDWR/projekt/, SDM vs SDM2, and EMISY/componentSchematic/ vs EMISY/schematic/ are likewise near-duplicates. They are not de-duplicated: each is a separate submitted deliverable, and merging them would rewrite the academic record.
• Programming/EOPSY/lab{3,4}/task*/work/ is not student code. It is the extraction target of make setup from the tarballs in the sibling ftp/. Verified byte-identical, 14/14 files. Do not edit or split it.

Git Workflow

Work directly on main; commit and push straight there.

./install.sh          # deps + pre-commit hooks, then verifies and fails loudly
./run.sh check        # the full local gate
./run.sh status       # probe every course dir, regenerate DOCS-runnability.md
./run.sh run <dir>    # build/run one course, capped at 2G RAM and 2 cores

run executes inside a transient systemd scope (MemoryMax=2G, MemorySwapMax=0, CPUQuota=200%, nice -n 19, ionice -c 3), so a course build cannot make an interactive machine stutter. This is not caution, it is a kernel-enforced ceiling -- PORR's suite solves 10000x10000 systems and will take every core it is given. Raise it deliberately for a single run with RUN_MEMORY_MAX=8G RUN_CPU_QUOTA=800% ./run.sh run <dir>.

Never --no-verify. If a gate fails, fix the cause.

Development Workflow

• Confirm a change works by running the thing and reading its output, not by running tests. Tests come after.
• git rm --cached, never bare git rm, when removing build output — it must stay on disk so a mis-classification is recoverable.
• After touching .gitignore or .binary-allowlist, re-run:

  git ls-files | git check-ignore --stdin --no-index
  

  It must print nothing. A file that is both tracked and ignored fails silently on git add: not an error, the edit just never lands. This repo had 130 such files.

Code Conventions

Anything a hook already enforces is omitted here — read .pre-commit-config.yaml for that list. What follows is what a linter cannot check.

Everything

• Every file is at most 250 lines, source and prose alike, enforced by the shared gate in ~/utils. There is deliberately no baseline and no allowlist: split the file. For a LaTeX report that means \input{} per section; for a markdown report, sibling DOCS-*.md files behind an index.
• Coursework is held to the same bar as the tooling. It gets the full select = ["ALL"] treatment -- annotations, docstrings, naming, the lot -- and there are no per-file-ignores carving the two course trees out. The version that was handed in is preserved by git history, which is a better archive than a frozen working tree; this repo's job is code that still builds, runs and reads well. (This reverses an earlier rule that said not to improve coursework beyond what was submitted. kuhy overruled it on 2026-08-28: "yes fix it if necessary we can always go back to worse version by going back in git repo.")
• The one thing that is still off limits is what was delivered: do not add content to a submitted .pdf/report that was never in it, and do not merge the near-duplicate submission directories listed above.

Vendored content

gates/vendored.txt is the only place a path becomes exempt from lint, coverage and the cap. Every entry carries a reason and how it was verified — an unverified entry is a suppression wearing a costume. The list is mirrored into ~/utils/file_length/_tables.py, [tool.ruff] exclude and [tool.coverage.run] omit; edit the manifest, never a mirror.

Python

• ruff runs with select = ["ALL"] and mypy with strict = true.
• No # noqa, no # type: ignore. A pre-commit hook rejects both outright.
• Notebooks: ruff reads and formats .ipynb natively (cell-aware, so an import used two cells later is not reported unused), which is why there is no nbqa layer. It rewrites source only — outputs are left byte-identical. Nothing strips outputs on commit yet; the four notebooks here were committed with theirs. Real logic belongs in a sibling importable module — a notebook cannot be covered.

LaTeX

Build output is ignored, not committed: .aux, .fls, .fdb_latexmk, .xdv, -eps-converted-to.pdf and friends. The submitted .pdf is committed, because it is the deliverable.

Shell

set -euo pipefail, quoted expansions, main "$@" last. Never embed another language's logic inline — put it in its own file so that language's tooling applies.

Key Files

Path	Why it matters
gates/vendored.txt	What is not my code, and how each entry was verified.
gates/coverage.tsv	Per-directory coverage floors. Floors may only rise.
.binary-allowlist	Path-scoped: deliverables allowed under the two course trees, compiled output allowed nowhere.
DOCS-runnability.md	Per-course run status, recorded separately for local and ci.
TODO-*.md	Open decisions. Each carries the removal marker and is deleted when the work lands.