Stabilize publication diagnostics retention test

[provencher]

Summary

• stop the live FSEvent stream during the publication-retention measurement window
• measure the 40 synthetic publications relative to the drained startup baseline
• assert coherent retained/dropped accounting and ordered synthetic samples

This fixes the full-suite CI race observed on release PR #135 (41/9 instead of the test's assumed 40/8) without changing production behavior.

Validation

• pair-agent diagnosis and cleanup pass
• focused test repeated after the fix
• full WorkspaceFileContextStoreTests suite
• full root test suite: 1,268 passed, 1 skipped
• strict Swift format/lint
• contribution commit and push preflights