Add an artifact-type suffix (_Talk / _Poster) to standardized filenames (#1404)

[jonfroehlich]

Closes #1404.

What & why

A downloaded artifact file should say what it is. Today a talk's exported slides, the paper, and the poster all generate the same basename, so Froehlich_MakingInTheHCIL_CHI2024.pdf in a Downloads folder doesn't reveal which one it is (and two of them collide into … (1).pdf).

Talks now end in _Talk, posters in _Poster. Publications and grants stay unsuffixed — a bare paper PDF is the default expectation, and that same generated name is the .bib download name (get_pub_filename).

Froehlich_MakingInTheHCIL_CHI2024.pdf         <- publication (unchanged)
Froehlich_MakingInTheHCIL_CHI2024_Talk.pdf    <- talk slides
Froehlich_MakingInTheHCIL_CHI2024_Poster.pdf  <- poster

The four decisions from the issue

1. Worth doing? Yes, scoped to the ambiguous types only.
2. Types + label: static Talk / Poster, from a FILENAME_TYPE_SUFFIX class attribute — not derived from talk_type. Poster has no type field, and talk_type is nullable and editor-editable, so a data-derived label would rename files on a metadata-only edit.
3. Placement: trailing. Everything before the segment stays byte-identical to the old scheme, so a paper and its talk still sort adjacently, and the -<timestamp> uniquifier appends cleanly after it. (The existing mid-name suffix slot is left alone.)
4. Retroactive, via restandardize_artifact_filenames — already an idempotent every-container-start step.

The subtle part: command ordering

backfill_original_filenames runs at entrypoint step 4.7b, before the re-standardization at 4.10b. Its "is this already standardized?" guard is what stops it recording a renamed file's current name as its Originally uploaded as provenance. The moment generate_filename changed, every already-renamed talk and poster would have failed that guard and had its old standardized name written in as a fake original — corpus-wide, on the very next deploy. The guard now accepts both schemes (generate_filename(..., include_type_suffix=False) yields the pre-#1404 name).

Also extracted: the name == base or name.startswith(base + "-") comparison that three commands had each copy-pasted is now one tested helper, matches_standardized_basename. And repair_diverged_artifact_filenames now also scans for orphans under the pre-#1404 base (that bug predates this scheme change).

Accepted tradeoff (verified against prod)

/media/publications/Bogus.pdf -> 404 from Django (serve_pdf: exact -> original_pdf_filename -> fuzzy)
/media/talks/Bogus.pdf        -> 404 from Apache  (never reaches Django)
/media/posters/Bogus.pdf      -> 404 from Apache

So external/indexed links to renamed talk and poster PDFs cannot be redirected from this repo. Publications are excluded from the suffix, so the most link-exposed, most-indexed subtree doesn't move at all. A follow-up issue asks UW CSE IT to let /media/talks/ and /media/posters/ fall through to Django; after that, a redirect is easy — the pre-#1404 name is still derivable from metadata, so nothing needs storing now.

Rollout — this PR is stage 1 of 3

Step 4.10b runs live on every container start, so without a gate the whole corpus would be renamed unattended on the next deploy.

1. This PR ships 4.10b with --dry-run. On -test, confirm debug.log lists only Talk/Poster rows and zero Publications.
2. Follow-up commit removes --dry-run -> -test performs the rename; verify a talk in /admin (pdf + raw + thumbnail suffixed, provenance row intact) and the public talk page's PDF link.
3. Tag 2.29.0 -> prod performs the one-time rename.

Verification

• Full suite: 661 OK (8 skipped).
• New tests: trailing segment placement, per-class labels (Talk/Poster suffixed; Publication/Grant not), matches_standardized_basename including that the two schemes don't match each other, retroactive rename of a pre-Discuss: add an artifact-type suffix (Talk/Poster/Demo) to standardized filenames #1404 talk (pdf + raw + thumbnail) plus idempotency on re-run, publications not churning, and the backfill guard for both a legacy name and its -<timestamp> variant.
• Real-data dry-run diff against the local dev DB (291 artifacts), old code vs. new: the rename set grows from 122 to 133 rows, and the delta is exactly 10 talks + 1 poster — the already-standardized ones — with no publications added and nothing dropped.

No model changes, no migration. No UI surface changes (all links render from pdf_file.url, so they self-heal), so no Pa11y run needed.

🤖 Generated with Claude Code