Audit: assert on lfs_media and on ref/path resolvability — strict exits 0 on a repoint that 404s every read

[mmcky]

The strict audit reports a media.githubusercontent.com read of a data-lectures path as fully migrated and exits 0. So the high_dim_data fold can be declared complete, migration.yml can be flipped to repointed, the dashboard stays green — and all six datasets are unreadable in every consuming lecture.

This is the first gate on QuantEcon/workspace-lectures#23 step 3. Its whole value is being red at the moment the lecture repoints hit main; landed after the fold it is worth nothing.

The hole

scripts/build_audit.py:397-398 derives migrated from the org/repo pattern alone, so media.githubusercontent.com/media/QuantEcon/data-lectures/main/lectures/<file> and raw.githubusercontent.com/QuantEcon/data-lectures/main/lectures/<file> are indistinguishable to every consistency check.

The flag that would separate them already exists. :196 computes "lfs_media": "media.githubusercontent.com" in url per reference, and the only thing that ever reads it is a dashboard badge at scripts/render_audit.py:841. Nothing asserts.

Reproduced end to end: a doctored lecture-python-intro and lecture-wasm with all 12 .md reads on the media host, migration.yml flipped to repointed, gives --strict EXIT=0, no warnings. That is the exact state a mechanical org/repo swap produces, which is what repoint rule 6 exists to prevent.

Worse, the comment introducing the rule-5 check at :462-463 currently reads "…must use raw.githubusercontent.com, or media.githubusercontent.com for LFS files." Since nothing under lectures/ is or will be an LFS object, that sentence now blesses the trap, and any patch has to amend it or leave the file self-contradictory.

The same shape, one step further out

branch_pinned at :195 is the identical computed-never-asserted pattern, read only by another badge at render_audit.py:843. And datasets are keyed by basename (:277, :388), so the path is never checked either.

Consequence: rewriting all 12 reads to raw.githubusercontent.com/QuantEcon/data-lectures/nonexistent-branch/WRONG_DIR/<file> also gives EXIT=0 — even with the lfs_media check applied. The host check alone is not enough.

A resolvability check is the better shape, and it subsumes the host one: for every reference with pattern == 'data-lectures', assert ref == 'main', path == 'lectures/<basename>', and that the file exists in this repo. That catches a wrong host, a wrong branch, a wrong directory, and a repoint that lands before its file — and it makes the Phase 4 data.quantecon.org cutover (#15) mechanically checkable later by flipping the expected form.

Zero false positives: all 33 real data-lectures references across the eight scanned repos today are exactly (main, lectures/<basename>). Verified: the check applied to today's main exits 0; applied to the doctored trees it exits 1 with 12 named warnings.

Notes for whoever writes it

• The code belongs after :473, the comment after :465 — they are not adjacent.
• It is roughly 12 lines, 7 of comment and 5 of code.
• :462-463 must be amended in the same change.
• Warning wording that reads well next to the existing rule-5 one: forbes-global2000.csv: lecture-python-intro heavy_tails reads it via media.githubusercontent.com — the LFS media host 404s plain-git files (repoint rule 6).

What this still will not catch

Worth stating so the gate is not oversold. The audit scans lectures/**.md only and excludes /_static/ (:215-216, :222-224), so the three _static/lecture_specific/inequality/data.ipynb:37 builder reads stay invisible — those are hand-edit-and-hand-check in all three repos. And SCAN_REPOS (:48-57) excludes translations by decision, so lecture-intro.zh-cn's seven reads of the same six files are outside the scan entirely. Both are noted in repoint rules 1 and 6 as amended in #53.

Making the .ipynb reads visible is a separate, small change — a ~5-line edit to the file filter plus the cell extractor surfaces both — and is worth doing on its own merits, but it is not this issue.

Rules and context: PLAN.md repoint rules 1, 5 and 6 (as corrected in #53). Plan: QuantEcon/workspace-lectures#23 step 3.