connector-google-calendar: drop orphan pending_occ leftovers

[KrisBraun]

Yesterday's full-pass-cleanup flush (a4ee2a8) saved every leftover pending_occ entry as a standalone occurrence-only link. Cross-batch cases (master in batch A, instance buffered in batch B) need this so the upsert attaches to the existing master, but when the master never appeared in any batch — i.e. the master event has been deleted from Google Calendar but a per-instance cancellation tombstone still surfaces in events.list — the flush turns into an INSERT of a brand-new link with title=undefined, no schedule, no notes. Empty Untitled threads materialize on the client, with no way to recover.

Fix: track masters seen across batches via seen_master:
storage markers (only during initial sync, since incremental sync doesn't buffer). At full-pass cleanup, only flush a leftover if its master canonical was processed during this pass; drop the rest. Markers are cleared at the end of cleanup. Stale markers from a crashed prior pass are harmless: if such a master is in the DB, the flush attaches correctly to it; if not, the seenMasters set rebuilds correctly from the new pass.

[KrisBraun]

Yesterday's full-pass-cleanup flush (a4ee2a8) saved every leftover pending_occ entry as a standalone occurrence-only link. Cross-batch cases (master in batch A, instance buffered in batch B) need this so the upsert attaches to the existing master, but when the master never appeared in any batch — i.e. the master event has been deleted from Google Calendar but a per-instance cancellation tombstone still surfaces in events.list — the flush turns into an INSERT of a brand-new link with title=undefined, no schedule, no notes. Empty Untitled threads materialize on the client, with no way to recover.

Fix: track masters seen across batches via seen_master:
storage markers (only during initial sync, since incremental sync doesn't buffer). At full-pass cleanup, only flush a leftover if its master canonical was processed during this pass; drop the rest. Markers are cleared at the end of cleanup. Stale markers from a crashed prior pass are harmless: if such a master is in the DB, the flush attaches correctly to it; if not, the seenMasters set rebuilds correctly from the new pass.

[KrisBraun]

Yesterday's full-pass-cleanup flush (a4ee2a8) saved every leftover pending_occ entry as a standalone occurrence-only link. Cross-batch cases (master in batch A, instance buffered in batch B) need this so the upsert attaches to the existing master, but when the master never appeared in any batch — i.e. the master event has been deleted from Google Calendar but a per-instance cancellation tombstone still surfaces in events.list — the flush turns into an INSERT of a brand-new link with title=undefined, no schedule, no notes. Empty Untitled threads materialize on the client, with no way to recover.

Fix: track masters seen across batches via seen_master:
storage markers (only during initial sync, since incremental sync doesn't buffer). At full-pass cleanup, only flush a leftover if its master canonical was processed during this pass; drop the rest. Markers are cleared at the end of cleanup. Stale markers from a crashed prior pass are harmless: if such a master is in the DB, the flush attaches correctly to it; if not, the seenMasters set rebuilds correctly from the new pass.

[KrisBraun]

Yesterday's full-pass-cleanup flush (a4ee2a8) saved every leftover pending_occ entry as a standalone occurrence-only link. Cross-batch cases (master in batch A, instance buffered in batch B) need this so the upsert attaches to the existing master, but when the master never appeared in any batch — i.e. the master event has been deleted from Google Calendar but a per-instance cancellation tombstone still surfaces in events.list — the flush turns into an INSERT of a brand-new link with title=undefined, no schedule, no notes. Empty Untitled threads materialize on the client, with no way to recover.

Fix: track masters seen across batches via seen_master:
storage markers (only during initial sync, since incremental sync doesn't buffer). At full-pass cleanup, only flush a leftover if its master canonical was processed during this pass; drop the rest. Markers are cleared at the end of cleanup. Stale markers from a crashed prior pass are harmless: if such a master is in the DB, the flush attaches correctly to it; if not, the seenMasters set rebuilds correctly from the new pass.

[KrisBraun]

Yesterday's full-pass-cleanup flush (a4ee2a8) saved every leftover pending_occ entry as a standalone occurrence-only link. Cross-batch cases (master in batch A, instance buffered in batch B) need this so the upsert attaches to the existing master, but when the master never appeared in any batch — i.e. the master event has been deleted from Google Calendar but a per-instance cancellation tombstone still surfaces in events.list — the flush turns into an INSERT of a brand-new link with title=undefined, no schedule, no notes. Empty Untitled threads materialize on the client, with no way to recover.

Fix: track masters seen across batches via seen_master:
storage markers (only during initial sync, since incremental sync doesn't buffer). At full-pass cleanup, only flush a leftover if its master canonical was processed during this pass; drop the rest. Markers are cleared at the end of cleanup. Stale markers from a crashed prior pass are harmless: if such a master is in the DB, the flush attaches correctly to it; if not, the seenMasters set rebuilds correctly from the new pass.

[KrisBraun]

Yesterday's full-pass-cleanup flush (a4ee2a8) saved every leftover pending_occ entry as a standalone occurrence-only link. Cross-batch cases (master in batch A, instance buffered in batch B) need this so the upsert attaches to the existing master, but when the master never appeared in any batch — i.e. the master event has been deleted from Google Calendar but a per-instance cancellation tombstone still surfaces in events.list — the flush turns into an INSERT of a brand-new link with title=undefined, no schedule, no notes. Empty Untitled threads materialize on the client, with no way to recover.

Fix: track masters seen across batches via seen_master:
storage markers (only during initial sync, since incremental sync doesn't buffer). At full-pass cleanup, only flush a leftover if its master canonical was processed during this pass; drop the rest. Markers are cleared at the end of cleanup. Stale markers from a crashed prior pass are harmless: if such a master is in the DB, the flush attaches correctly to it; if not, the seenMasters set rebuilds correctly from the new pass.

[KrisBraun]

Yesterday's full-pass-cleanup flush (a4ee2a8) saved every leftover pending_occ entry as a standalone occurrence-only link. Cross-batch cases (master in batch A, instance buffered in batch B) need this so the upsert attaches to the existing master, but when the master never appeared in any batch — i.e. the master event has been deleted from Google Calendar but a per-instance cancellation tombstone still surfaces in events.list — the flush turns into an INSERT of a brand-new link with title=undefined, no schedule, no notes. Empty Untitled threads materialize on the client, with no way to recover.

Fix: track masters seen across batches via seen_master:
storage markers (only during initial sync, since incremental sync doesn't buffer). At full-pass cleanup, only flush a leftover if its master canonical was processed during this pass; drop the rest. Markers are cleared at the end of cleanup. Stale markers from a crashed prior pass are harmless: if such a master is in the DB, the flush attaches correctly to it; if not, the seenMasters set rebuilds correctly from the new pass.

[KrisBraun]

Yesterday's full-pass-cleanup flush (a4ee2a8) saved every leftover pending_occ entry as a standalone occurrence-only link. Cross-batch cases (master in batch A, instance buffered in batch B) need this so the upsert attaches to the existing master, but when the master never appeared in any batch — i.e. the master event has been deleted from Google Calendar but a per-instance cancellation tombstone still surfaces in events.list — the flush turns into an INSERT of a brand-new link with title=undefined, no schedule, no notes. Empty Untitled threads materialize on the client, with no way to recover.

Fix: track masters seen across batches via seen_master:
storage markers (only during initial sync, since incremental sync doesn't buffer). At full-pass cleanup, only flush a leftover if its master canonical was processed during this pass; drop the rest. Markers are cleared at the end of cleanup. Stale markers from a crashed prior pass are harmless: if such a master is in the DB, the flush attaches correctly to it; if not, the seenMasters set rebuilds correctly from the new pass.