Remove the hardcoded data paths from the scripts and their tests

[emrefbulut]

What changed

Four files carried an absolute path into one developer's temporary directory as a fallback:

• scripts/leakage_real.py — DASH7 source
• scripts/leakage_loraiq.py — LoRaIQ source + three CSVs
• tests/test_preflight.py — LoRaIQ resolver
• tests/test_measurement.py — the same resolver, duplicated

All four now read IQFORGE_DASH7 / IQFORGE_LORAIQ (plus optional IQFORGE_LORAIQ_INDEX, _LABELS, _GROUPS) with no fallback. Every path keeps the command-line flag it already had, and the flags still win.

The duplicated fifteen-line resolver now lives once in tests/helpers.py.

docs/methodology.md §Reproducing shows the variables. tests/test_repo_hygiene.py is new.

Why

The path resolved on exactly one machine. The scripts ran there and were unrunnable everywhere else; the LoRaIQ tests passed there and skipped everywhere else — while docs/methodology.md opens with "Everything below is reproducible from this repository."

The fallback is the part that made it invisible. Without one, an absent dataset is an error the first person to run the script sees immediately. With one, the author never sees anything at all, and everyone else sees a skip that reads like a fact about their machine rather than about the repository.

The skip reason mattered too. "LoRaIQ recordings are not on this machine" cannot be acted on: it does not distinguish you never configured this from you configured it and mistyped the path, and those need opposite fixes. Now:

SKIPPED [1] tests/test_preflight.py:237: IQFORGE_LORAIQ is not set. Point it at a
directory of prepared LoRaIQ recordings to run this; the dataset is a public
download and is not in this repository.

and for a set-but-wrong path, IQFORGE_LORAIQ is set to '...', which is not a directory.

How it was verified

Control — the tests genuinely run when the variable is set, not just skip more politely:

IQFORGE_LORAIQ=<real path> uv run pytest tests/test_preflight.py -k loraiq_recordings_are_not_refused -q
# 1 passed

Three mutations, each red on two tests — the mechanical scan and the behavioural one:

mutation	caught by
reinstate the original fallback in the resolver	test_no_source_file_hardcodes_a_machine_path[tests/helpers.py] + test_the_resolver_finds_nothing_when_the_variable_is_unset
make the skip reason generic again	test_the_loraiq_skip_reason_names_the_variable_to_set + test_the_resolver_finds_nothing_when_the_variable_is_unset
give leakage_loraiq.py a default path again	test_no_source_file_hardcodes_a_machine_path[scripts/leakage_loraiq.py] + test_the_experiment_scripts_expose_no_default_data_path

The first mutation is the original defect, restored verbatim. It is caught.

uv run pytest tests/ -q --ignore=tests/test_preflight.py   # 342 passed, 1 skipped
uv run pytest tests/test_preflight.py -q                   # 16 passed, 1 skipped
uv run ruff check .                                        # clean

The new hygiene test

Convention 8 is the one with no natural failure mode — the mistake works perfectly for whoever makes it, which is why it survived four files and a documented claim. So the check is mechanical: tests/test_repo_hygiene.py scans src/, scripts/ and tests/ for string literals anchored at a Windows drive letter or a POSIX home/tmp root, one parametrised case per file so the failure names the file.

A line that only looks like a machine path opts out with an inline not-a-machine-path marker rather than a file-level allowlist — the exemption then sits where a reviewer reads it, next to the string it excuses. Exactly one line qualifies today: synthetic overflow data in the audit width test ("C:/a-very/deep/path/" * 8).

Note for the reviewer

Two files beyond the ones named in the task had the identical defect: tests/test_measurement.py (same duplicated resolver) and scripts/leakage_real.py (DASH7 source). Fixing only the two named ones would have left the hygiene test failing on the other two, and left --sweep stride on DASH7 unrunnable. Say if you would rather split them out.

Conventions it touches

• 1 Never fall back silently — this PR is that rule applied to configuration. The fallback path is exactly a silent fallback: it produced a plausible working state instead of a refusal.
• 2 A passing test does not prove it can fail — three mutations, including the original defect restored verbatim. Plus a control proving the tests run rather than skip when configured.
• 8 Nothing session- or machine-specific is committed — now enforced by a test instead of by review.