Coalesces edits that arrive during view publication into one follow-up per root. The running publication completes, and subsequent edits accumulate as a single pending batch computed against the newly committed manifest instead of cancel-and-restart.
Behavior
Edits no longer cancel an in-flight publication; root deletion, unbind, CAS conflict, and configuration changes still do.
Failed batches are retried with both active and pending paths; a rescan follow-up is promoted to a full read exactly once, and incompatible configurations force a full read.
Adds differential tests for edit bursts, failed batches, rescans, unbind, configuration changes, and competing winning publications, each verified against a cold full-tree manifest.
Confidence Score: 5/5
The PR appears safe to merge; no actionable issues were found.
Summary
View publication now keeps one worker per root and combines incoming edits into a pending batch.
View edits queue behind one publication per checkout.
No actionable issues were found. Tests were inspected but not run.
Diagram
%%{init: {'theme': 'neutral'}}%%
flowchart TD
A[Schedule publication] --> B{Worker exists for root?}
B -->|No| C[Start worker]
B -->|Yes| D[Merge or replace pending batch]
C --> E[Prepare and publish current batch]
E --> F{Attempt succeeded?}
F -->|Yes| G{Pending batch exists?}
F -->|No| H[Retain failed paths for retry]
H --> I[Merge compatible pending work]
I --> E
G -->|Yes| J[Take pending batch]
J --> E
G -->|No| K[Remove worker under scheduler lock]
