v26.05.25.1: maintenance sweep + Camie v2 + corrupt-file handling + post-date gallery + clear-stuck escape hatch #11

2026-05-25T12:57:09-04:00

bvandeusen commented

2026-05-25 12:57:09 -04:00

(Migrated from git.fabledsword.com)

Patch release rolling up 10 fixes from the post-v26.05.25.0 work session.

Maintenance / recovery

9d5abb0 recover_interrupted_tasks now sweeps pending/queued orphans (>30 min) to failed, not just processing (>5 min). Fixes the 5,490-row pile operator hit when scan_directory crashed mid-enqueue.
b6a917a /api/import/clear-stuck endpoint + "Clear stuck..." UI button on the import task list. Manual escape hatch for the autoretry-loop case the automatic sweep can't break (e.g., corrupt file PIL OSError loop).
3a359f6 Quick scan button always visible, disabled when an active batch is present, with an inline "Clear stuck" action next to the spinner. Operator was unable to start scans because the wedged batch hid the button.

ML pipeline

3b3e756 Tagger + downloader aligned with actual Camie v2 HF layout: camie-tagger-v2.onnx (not model.onnx), JSON metadata (not selected_tags.csv), ImageNet normalize, NCHW, sigmoid on refined logits. Verified against the published onnx_inference.py reference.
6b1bb87 Pinned-test rename to match Camie v2 filenames.

Importer resilience

68cffce Importer catches PIL.OSError during _transparency_pct + compute_phash and skips as invalid_image instead of letting Celery autoretry loop forever on a corrupt file. Operator hit this with a broken JPEG in the IR set.

Gallery sort/display

c361032 Gallery scroll / timeline / jump / neighbor-nav now sort/group by COALESCE(post.post_date, image_record.created_at). Migrated content surfaces at its original publish date, not FC's scan-week date. tag_apply phase 4 now sets image_record.primary_post_id (idempotent, doesn't clobber importer-set values). API response gains posted_at + effective_date fields alongside the existing created_at for future modal display.
9f54efd tag_apply phase 4 now covers deviantart + pixiv (was silently dropping ALL PostMetadata for those platforms). Operator's missing IR provenance had this as one of two root causes.

UI polish

52445eb Directory card width 220px → 440px; preview slots get explicit min/max-height + overflow: hidden so tall source images can't escape on browsers where aspect-ratio doesn't compute.

Migration recovery notes

After deploy, to backfill IR-migrated images with their post dates and pick up the previously-skipped DeviantArt + Pixiv entries:

Run /api/migrate/tag_apply again on the existing manifest (phase 4 is idempotent; existing rows pass through; new primary_post_id and new platform-coverage entries backfill).
Gallery + timeline + jump immediately reflect publish dates after that.

Schema changes

None. All fixes are logic-only.

🤖 Generated with Claude Code

Patch release rolling up 10 fixes from the post-v26.05.25.0 work session. ## Maintenance / recovery - **`9d5abb0`** `recover_interrupted_tasks` now sweeps pending/queued orphans (>30 min) to `failed`, not just `processing` (>5 min). Fixes the 5,490-row pile operator hit when `scan_directory` crashed mid-enqueue. - **`b6a917a`** `/api/import/clear-stuck` endpoint + "Clear stuck..." UI button on the import task list. Manual escape hatch for the autoretry-loop case the automatic sweep can't break (e.g., corrupt file PIL OSError loop). - **`3a359f6`** Quick scan button always visible, disabled when an active batch is present, with an inline "Clear stuck" action next to the spinner. Operator was unable to start scans because the wedged batch hid the button. ## ML pipeline - **`3b3e756`** Tagger + downloader aligned with actual Camie v2 HF layout: `camie-tagger-v2.onnx` (not `model.onnx`), JSON metadata (not `selected_tags.csv`), ImageNet normalize, NCHW, sigmoid on refined logits. Verified against the published `onnx_inference.py` reference. - **`6b1bb87`** Pinned-test rename to match Camie v2 filenames. ## Importer resilience - **`68cffce`** Importer catches `PIL.OSError` during `_transparency_pct` + `compute_phash` and skips as `invalid_image` instead of letting Celery autoretry loop forever on a corrupt file. Operator hit this with a broken JPEG in the IR set. ## Gallery sort/display - **`c361032`** Gallery scroll / timeline / jump / neighbor-nav now sort/group by `COALESCE(post.post_date, image_record.created_at)`. Migrated content surfaces at its original publish date, not FC's scan-week date. tag_apply phase 4 now sets `image_record.primary_post_id` (idempotent, doesn't clobber importer-set values). API response gains `posted_at` + `effective_date` fields alongside the existing `created_at` for future modal display. - **`9f54efd`** tag_apply phase 4 now covers `deviantart` + `pixiv` (was silently dropping ALL PostMetadata for those platforms). Operator's missing IR provenance had this as one of two root causes. ## UI polish - **`52445eb`** Directory card width 220px → 440px; preview slots get explicit min/max-height + `overflow: hidden` so tall source images can't escape on browsers where `aspect-ratio` doesn't compute. ## Migration recovery notes After deploy, to backfill IR-migrated images with their post dates and pick up the previously-skipped DeviantArt + Pixiv entries: 1. Run `/api/migrate/tag_apply` again on the existing manifest (phase 4 is idempotent; existing rows pass through; new `primary_post_id` and new platform-coverage entries backfill). 2. Gallery + timeline + jump immediately reflect publish dates after that. ## Schema changes None. All fixes are logic-only. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: bvandeusen/FabledCurator#11