115 ruff findings to 0. All three notebooks' non-source content hashes the
same before and after -- c7660e3e5e3b8d4ffc008d7f, b918553ae0cf91cd56fa4091
and 3f09c77e3edae1727e3781e2 -- so every stored output, execution count and
piece of metadata is exactly what was submitted.
Most of the work was 35 over-long lines, and they are over-long because they
are full Polish GUS column descriptions used as dict keys in .rename(). A key
like "udziały w podatkach stanowiących dochody budżetu państwa podatek
dochodowy od osób fizycznych" cannot be shortened, only wrapped, so a script
re-wrapped all of them with implicit concatenation inside parentheses -- the
one form both ruff's linter and its formatter accept.
A script doing 134 mechanical edits to strings that are database keys is
exactly the kind of change that silently drops a space. So it is checked
mechanically: an AST pass collects every string dict key in every code cell of
both versions and compares the multisets. 120 keys in data.ipynb, 9 in
main.ipynb, 5 in pbad.ipynb, all identical.
The other 27 findings were commented-out code (ERA001), and deleting it would
have thrown away what it recorded: which other operational programmes were
run, which axis limits the plots were read at, that MinMaxScaler was tried and
made no difference, that an earlier pass dropped a hand-maintained column list.
Each is now a sentence of prose saying that, and git history still has the
code. Two cells were nothing but commented-out scratch and are now just the
note.
Also real: five `inplace=True` calls became assignments, `zip()` in three
feature-importance loops got `strict=True`, `max(max(a), max(b))` became
`max(*a, *b)`, and 1.5 and 7 became DENSE_GMINA_THRESHOLD and
TERYT_GMINA_DIGITS -- the latter being "a full TERYT gmina code is seven
digits", which is worth saying once rather than leaving as a bare literal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
228 ruff findings to 0 across the notebook, its partial export and twm.py, and
mypy --strict is clean on both .py files.
A notebook is a deliverable that happens to contain code, so the check that
mattered was not "does it still run" -- it needs a GPU, MNIST, CIFAR-10 and a
slice of Quick, Draw! to run at all -- but "is anything except the code still
exactly as submitted". Hashing every key of the .ipynb EXCEPT cell `source`
gives 60242e116d12549d11c93f7f before the first edit and after the last one,
across four passes and a `ruff format`. The file also round-trips byte-for-byte
through json.dumps(indent=1, ensure_ascii=False), which is what made editing
cells programmatically safe rather than a reformat of all 141 of them; the
diff touches 38.
print() became sys.stdout.write() rather than logging. In a notebook the print
IS the output, and a logger with no handler configured would have silently
emptied every cell -- the exact regression scripts/check_refactor_safety.py
was written for. matprint's `print(x, end=" ")` and bare `print()` are
preserved as writes with and without a trailing newline.
Both `show_samples` and `show_samples_rgb` took (indices, preds, images,
labels, count, names), which is six arguments and a mutable default. They now
take a frozen `Samples` holding the last four, which also removed the same
four-value tuple from seven call sites. Reviewing that diff is how I caught
that I had transcribed `real_score = preds[sample][real_id]` as
`samples.labels[sample]` -- a bug I introduced and then fixed before commit.
Other real things, not just rule satisfaction: `np.in1d` is removed in NumPy 2
(NPY201) and is now `np.isin(...).ravel()`; `test_model` computed
correct_indices and incorrect_indices and used neither; two cells grew `img`
inside a loop so it was only bound by the `else` branch; and the Quick, Draw!
downloader now builds a URL whose https scheme is a literal, which is what
S310 was asking for.
The README claimed the .py was "an nbconvert export of that notebook". It is
an export of the CIFAR-10 section only -- 28 of 81 code cells -- and has been
edited since. It says that now, and says the two are separate artifacts rather
than one being regenerable from the other.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
494 ruff findings to 0, and mypy --strict is clean too -- the first course in
this campaign where that is true of both.
Two real bugs turned up on the way, neither of which ruff can see:
- `main.py` passed `args.environment_map` to `ray_trace`, and argparse only
ever defined `--environment`. Every `--algorithm ray_tracing0` run died with
AttributeError before rendering a pixel, so `rendering.py` -- the largest
file in the course -- had no way to be exercised at all.
- `mouse_button` in the environment-map viewer assigned `mouse_last_x/y`
without declaring them `global`, so the drag origin went to locals and was
thrown away. Measured against HEAD: a 10x5 px drag turned the camera by
11.0/20.5 degrees because it measured from (0, 0); it is 1.0/0.5 now, which
is what the 0.1 degrees-per-pixel constant says it should be.
Fixing the first one is what made the rest safe. `ray_tracing0` has no RNG
anywhere in it, so it renders byte-identically every time, and the whole
refactor of `rendering.py` is verified by one md5 comparison:
02484bdb6b6c1ddeff317d7e5566580f, 248619 rays, before and after. The
cornell_box path cannot do that -- its ray count drifts by ~20 between
identical runs.
Three files were over the 250-line cap and are now split, geometry away from
the algorithm that uses it: rendering.py 462 -> 137 + raytracer.py,
photon_mapping.py 261 -> 160 + photon_geometry.py, and
code/photonmapping/python/main.py 257 -> 203 + geometry.py. No TRAK source
file is over the cap now.
The two photon mappers do use numpy's legacy global RNG, so NPY002 could not
be bit-identical. Each was done in two steps: restructure first, verified
identical under an external seed (digests 7220370f60206a62 and
21b4771504ab2351, and an identical 828 photon traces), then the swap to
`default_rng` on its own, checked against the before/after distributions.
Verified by running, not by reading:
- cornell_box through `./run.sh run Programming/TRAK`: 24.27s, 13294 rays
- all five scenes build after `from sightpy import *` became explicit imports
- the standalone photon mapper end to end: 81.4s, 38742 photon traces
- the EXR loader byte for byte: a0c3890350ce6101247c6948, 6291456 bytes
- the GLUT viewer under xvfb-run, which drew a real 1024x512 frame
`pyproject.toml` gains a mypy override for Imath, OpenEXR, OpenGL and
matplotlib. Those four ship neither stubs nor py.typed, and without it mypy
cannot import them at all and so reports nothing else about the file.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
1,183 ruff findings left, in five courses. Nine are at zero. The numbers in
here were four hours stale.
Two things worth reading before the next pass: the notebook branch of
scripts/course_run.sh will overwrite tracked notebook outputs the moment it is
used (the same bug already found in TRAK and SPD), and eleven .py files are
now over the 250-line cap -- several because THIS work added the docstrings
and annotations that pushed them over.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172dj34FDpao6wsix8ZuCzE
Running the courses made this machine stutter during a game. That was me:
PORR's suite solves systems up to 10000x10000 and takes every core it is
given, and the pip installs are not cheap either.
A promise to be careful is not a fix, so `run` now executes inside a transient
systemd scope -- MemoryMax=2G, MemorySwapMax=0 (OOM-kill rather than thrash
swap), CPUQuota=200%, nice -n 19, ionice -c 3. Same mechanism the pre-push
hook in testsAndMisc uses. Verified from inside the scope: memory.max reads
2147483648 and cpu.max reads "200000 100000", so the kernel is enforcing it
rather than the script hoping.
RUN_MEMORY_MAX and RUN_CPU_QUOTA override it for a single deliberate run.
Recorded the two courses that still fail: PORR, which will now be OOM-killed
under the cap and needs a smaller default parameter set, and ZPOB, where
latexmk exits 12 from bibtex despite producing the PDF.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172dj34FDpao6wsix8ZuCzE
The remaining work is much larger than what has landed, and vague about it
would be worse than useless. Numbers are measured, not guessed:
- 9,281 ruff findings under the fleet config (select = ALL). ~3,000 are
auto-fixable and ~6,000 are hand work, dominated by 1,671 missing type
annotations and ~600 missing docstrings.
- 19 tests for ~28k lines of authored code, across 12 languages.
- 73 files still over the 250-line cap, 27 of which are blocked on the
decision parked in TODO-file-length-250.md.
- 22 dependency manifests, none pinned.
Also records the ordering rules that are easy to lose: the dependency upgrade
and its gate must land in the same commit; source files must not be split
before their tests exist; and the file-rewriting hooks are held back so that
touching hundreds of submitted coursework files is its own reviewed change.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0172dj34FDpao6wsix8ZuCzE
