[jeremydmiller]

Fixes #519

Everything else in #519 is already on main (#520 compliance additions + suite inventory, #523 tombstoned natural keys). This PR settles the last seam, AsyncDaemonCompliance.an_in_flight_append_below_committed_events_is_not_stepped_over.

It cannot be implemented honestly on SQLite, so SupportsHoldingAnAppendOpen stays false. The fact holds an append open on stream A, then commits stream B through the same store, and requires A's sequence to be allocated-but-uncommitted below B's.

• Allocating the sequence is a write, and a write holds the file's one write lock until commit, so B's commit blocks behind A (measured earlier: 3 min, then lock failure).
• WAL still admits one writer. DEFERRED vs IMMEDIATE only changes when the lock is taken (at the first write, the one we need held).
• A second database file would make the held row invisible to the store/daemon under test, so the fact would assert nothing.

The state the fact guards against cannot exist here (see FisherHighWaterDetector, high_water_contiguity_audit). Change is a comment on the fixture recording this; no behaviour change. Fisher.Tests green on net10.0 and net9.0 (2902 passed, 23 skipped each); enrolled counts unchanged.

🤖 Generated with Claude Code

[jeremydmiller]

Note for #519: the flag is left off for the reason above, which #519's done criteria allow.