summaryrefslogtreecommitdiffstats
path: root/tmp-decisions.md
diff options
context:
space:
mode:
authorChristophe Besson <cbesson@gmail.com>2026-08-25 00:29:24 +0200
committerChristophe Besson <cbesson@gmail.com>2026-08-25 00:29:24 +0200
commitd427118bd91d67f1a041e5daf267aebcd34ca9d7 (patch)
treeff9e0e71d4b115a871cf0902739942e6214b27d0 /tmp-decisions.md
parent06c101154d544290d27f18d6fc08fb5f58a4e5d5 (diff)
downloadmeshbay-0.7.tar.gz
fix(node): scope _enriched_attempted by group, not just content hash0.7
Major finding: entry.id is a content hash, so the exact same physical file — the same MP3, byte-for-byte — indexed into two different groups (a shared library reused across several demo/test groups, or genuinely the same folder shared into two groups) produces the *same id* in both. _enriched_attempted was a single flat set of bare ids shared across every group this node hosts. The moment one group's copy got enriched, every other group's otherwise-identical copy read as "already attempted" and was skipped forever — nothing else ever revisits an id once it's in this set. That group's Music tab (or Videos tab, same bug, same set) showed every affected file at duration 0 with no artist/album/thumbnail, permanently, no matter how long you waited or how many times you reloaded — group A having been enriched first was enough to silently starve every later group of the same content. Now keyed by (group_id, entry.id) throughout — the enrichment gate, the sweep, and the rename re-enrichment path, for both video and audio (they already shared the one set, and the collision risk is identical for both). New regression test constructs two groups with byte-identical audio content and confirms both enrich independently.
Diffstat (limited to 'tmp-decisions.md')
0 files changed, 0 insertions, 0 deletions