Retrying an upload must rebuild stale indexes

scope: generic · severity: trap · confidence: proven · subsystem: build

Evidence

https://github.com/porthole-dev/pmaports/commit/461e2ed; https://github.com/porthole-dev/pmaports/actions/runs/36668882264; https://github.com/porthole-dev/pmaports/actions/runs/36668882249

Symptom — A publication retry finds identical package assets and reports nothing new, but the signed index still omits those packages.

Cause — Uploading packages and uploading the index are separate operations. If publication stops between them, identical assets do not establish that the index was updated. A publisher that ignores repository download failures can also sign a partial inventory and lose older packages from the index.

What to do — Compare names in the verified signed index with the complete release APK inventory. Force reindexing on a mismatch, even if payload bytes are unchanged. Refuse failed or incomplete downloads. Use the same publisher for every release lane, and compare downloaded assets before declaring success.

Control — The production index/assets comparison runs against three local fixtures: a matching index, an orphaned new APK, and an empty host repository. Only the orphaned case requests reindexing. The workflow regression rejects Chromium’s old separate publisher; the shared path and its CI pass. These controls establish retry behavior, not an observation that the specific interrupted-upload fixture happened in production.