Fix flush for initial writes

[mark0-cn]

Summary

• track operations written directly before the first publish
• make has_pending_operations report those initial writes
• make flush publish initial writes as documented
• preserve empty extend as a non-pending operation

Root cause

Before the first publish, operations are applied directly to the write-side copy as an initialization optimization and never enter the oplog. has_pending_operations only inspected the oplog boundary, so it returned false after an initial write. flush consequently skipped the publish and readers continued to see the old value.

The new first_pending bit records whether that direct-write path consumed any operations and is cleared by the first successful swap.

Validation

• cargo test --offline
• cargo fmt --all -- --check
• regression coverage for empty extend and append -> flush -> read