Fix the entry points left unadjusted from the template

[peterbjohnson]

Closes #14

Four defects inherited from the boilerplate. None affected the deployed eval path, but each misled anyone working on the repo.

Approach

Tests first, in dev_test.py (new, following the existing <module>_test.py convention) and a rewritten preview_test.py. Before the change:

12 failed, 1 passed

The one pass was the old preview test, which asserted only that a preview key existed and was not None. That is why the sympy echo survived this long.

1. dev.py failed immediately

The command the README pointed at:

python -m evaluation_function.dev "<response>" "<answer>"
AttributeError: 'dict' object has no attribute 'to_dict'

Two faults. It called .to_dict() on a value that is a plain dict, and it passed its arguments as (answer, response) while the signature is (response, answer). The argument swap is the nastier one: it does not raise, it just silently exchanges "missing" and "extra" in the feedback. It now prints the result as JSON and passes the arguments the right way round.

2. preview.py was the sympy placeholder

It echoed the response back under a sympy key, which is meaningless for MIDI. It now reports what the system actually read:

14 notes, 8.0 s, C4 to A4.

An audio path is named but deliberately not transcribed, because the preview runs while the student is still working and transcription takes seconds:

Audio recording practice.wav, transcribed on submission.

The platform's Preview type carries only sympy and feedback, so the summary goes in feedback.

I implemented this rather than removing the preview registration, which #12 offered as the alternative. Removing it would mean command: preview starts returning an error, and I would not want to change what the platform sees on the strength of a maintenance pass. Say the word if you would rather it went.

3. Wrong return annotation in evaluation.py

Annotated -> Result, referring to lf_toolkit's Result class, which this function has never returned. Worth noting this is not merely cosmetic: actually returning a Result would change the output, because that class renders feedback by joining items with <br>, which would mangle the newline-separated message this function produces. The dict is correct, so the annotation is what changes. Three tests in section 8 now pin the returned shape.

4. Healthcheck failed inside the container

command: healthcheck runs the test suite, and evaluation_test.py reads its bulk cases from data/, which the image did not carry:

FileNotFoundError: /app/data/longMIDIsequence.json

.dockerignore excluded the whole data/ directory, so simply adding a COPY failed the build outright. It now admits that one fixture and nothing else, so datasets and recordings that land in data/ still stay out of the image. The fixture load also tolerates the file being absent, rather than failing the whole module at import and taking every other test down with it.

Verified in a container built from this branch, with #12's backend fix layered on so it would boot on arm64:

	Before	After
tests_passed	false	true
errors	evaluation_function.evaluation_test	none
successes	1	70

Only the 56 KB fixture lands in /app/data.

Verification

Check	Result
New tests, before	12 failed, 1 passed
New tests, after	13 passed
Full suite	88 passed
CI lint gate (E9,F63,F7,F82)	0
Container healthcheck	70 passed, 0 errors

Note on test file placement

dev_test.py is a new file, but it mirrors dev.py exactly as preview_test.py mirrors preview.py, so it follows the convention already in the repo rather than adding another ad-hoc one. The return-shape tests went into section 8 of evaluation_test.py, where the other evaluation_function integration tests live, deliberately avoiding a new numbered section that would collide with #13's section 11.

🤖 Generated with Claude Code

[ada-jz7125]

There are text assertions in preview_test.py, but I think this keyword assertion is acceptable as long as the preview is not changed completely