Bonus dive

Lessons learned

Preserve patch text literally in orchestration scripts

Expected: a single orchestration call would apply and commit four independent Python files in sequence.
Actual: Python docstrings containing Markdown backticks terminated the JavaScript template literal that held each patch, so evaluation failed before any file was created.
Next time: send patches containing arbitrary prose directly to apply_patch, or escape template delimiters before embedding them in an orchestration script. Confirm repository status after any orchestration failure before retrying.

Expected: the sibling repositories' documented python command would run the first local test checkpoint.
Actual: this shell exposes python3 but no python alias, so the verification command failed before test discovery.
Next time: use python3 before the course virtual environment exists, then use .venv/bin/python after setup. Keep learner-facing commands as python because activating the documented virtual environment provides that alias.

Expected: python3 -m unittest discover -v would recurse into the newly created test directory.
Actual: discovery reported zero tests because tests/ was not importable; a zero-test command still exited successfully.
Next time: create tests/__init__.py before the first test file and always wrap discovery with an explicit countTestCases() > 0 assertion in CI.

Expected: nearest-rank p95 over values 0 through 19 would select 19.
Actual: ceil(0.95 * 20) is one-based rank 19, which is zero-based index 18; the implementation was correct and the hand-written test expectation was not.
Next time: write out n, the one-based rank, the zero-based index, and the selected value when reviewing percentile examples. Include non-one-origin data so rank and value cannot be conflated.

Expected: duplicate normalized package names in pylock.toml would always be ambiguous and should fail the teaching audit.
Actual: the PyPA specification permits multiple entries for one package when their markers or sources narrow to one installable entry.
Next time: distinguish a format's legal representational variants from exact duplicate records. State what the teaching implementation does not resolve, but do not reject valid standard behavior merely to simplify it.

Expected: shifting every load-test timestamp by one hour would leave the duration-derived measurements identical, so the example could label the shifted summary "same measurements".
Actual: adding 3,600 to each timestamp perturbs the low-order bits. The printed values differ at the thirteenth digit, so the lesson asserted an equality the reader could see was false. The test was already honest, using assertAlmostEqual; only the learner-facing surface overstated the result.
Next time: when a metamorphic relation runs on floats, print the observed drift and the tolerance it is compared against. Reserve == for exact quantities such as ratios of counts, and say which of the two a given field is.

Expected: a fault table indexed by call number would exercise the writer's after-commit failure on every attempt it was scripted for.
Actual: the idempotent deduplication path returned the cached receipt before reaching the after-commit check. It still advanced the call index, so scripting two lost responses in a row silently discarded the second one and the retry reported success. No test caught it because every test scripted a single fault.
Next time: if a stimulus table advances on every call, apply it on every branch that call can take, or make the unreachable branches reject the stimulus loudly. Test each fault at length two, not only length one; the second occurrence is what exposes a branch that consumes without acting.

Expected: a suite with a test file per module and adversarial capstone scenarios would already cover the behavior the textbook describes.
Actual: seeding fifteen deliberate defects showed thirteen caught and two survivors. Both survivors were behaviors the textbook states outright: unknown scanner severities failing closed, and side effects being counted per invocation rather than per writer. Both were implemented correctly and simply never asserted.
Next time: treat a sentence in the textbook as a test obligation. Before declaring a lesson finished, mutate the branch that sentence describes and confirm something red appears.