[jeremydmiller]

Part of #519. Everything is done except one seam, left off on purpose (below).

Merged with main after #521 (1.28.0 release prep) and #523 (the natural-key fix) landed, so this PR is now only the compliance additions.

Done

• The natural-key tombstone fix comes from Tombstoning a stream frees its natural key (fisher#522) #523 now. NaturalKeyCompliance's two tombstone facts failed with DuplicateNaturalKeyException, because TombstoneStreamOperation left the natural-key lookup rows behind (they have no FK). This PR had its own fix for that. Tombstoning a stream frees its natural key (fisher#522) #523 landed the same fix on main (same column choice, same tenant term under conjoined tenancy, same set of tables), so the merge kept main's and dropped this PR's copy, including its EventGraph.QuotedNaturalKeyTableNames() helper, which nothing uses any more.
• Pins and version follow main (Fisher 1.28.0: release prep #521: Fisher 1.28.0, JasperFx.* 2.100.0, Weasel.* 9.49.0). The one pin this PR still changes is smoke/aot-consumer/AotConsumer.csproj: JasperFx.Events.SourceGenerator goes from 2.98.0 to 2.100.0, to match Directory.Packages.props (main had not bumped it).
• SupportsSessionTenantSwitching => true with a ForTenant override. Fisher's session is not an ITenantedQuerySession<>, and session.ForTenant returns a real FisherSession. The in-one-save fact passes.
• compliance_suite_inventory: ComplianceSuiteInventory.AssertEnrolled for the event fixture (the only excuse is SingleTenantedEventSlicingCompliance, jasperfx#727) and for the document fixture. The class lives in namespace Fisher.Tests, not Fisher.Tests.Compliance, on purpose: scripts/check_scoreboard.py counts every test class in that namespace as an enrolled suite. In that namespace this audit would read as a 64th enrolled suite of the 64 that 2.100.0 ships, while it exists to prove that exactly one of them is not enrolled. The compliance scoreboard therefore stays at 63 suites and 783 tests. HANDOFF's header (2996 total, 2925 in Fisher.Tests) and README's Fisher.Tests count take the two new tests.

Left out

• SupportsHoldingAnAppendOpen stays false. I implemented HoldAppendOpenAsync (an enlisted session on a BEGIN IMMEDIATE) and the fact failed after 3 minutes. SQLite allows one writer per file, so the held write transaction blocks the later committed append the fact needs. The state the fact needs, an uncommitted lower sequence under a committed higher one, cannot exist on SQLite. The reason is written at the flag, and high_water_contiguity_audit pins the property.

Local verification after the merge (CI is off)

dotnet build fisher.slnx: 0 errors. ./build.sh CI --framework net10.0 and --framework net9.0 both pass, and python3 scripts/check_scoreboard.py --configuration Release agrees with both runs: 2996 tests, 783 compliance across 63 suites (49 event / 14 document).

Command	Passed	Skipped	Failed
dotnet test src/Fisher.Tests/Fisher.Tests.csproj -f net10.0	2902	23	0
dotnet test src/Fisher.Tests/Fisher.Tests.csproj -f net9.0	2902	23	0

The same counts on both frameworks, by class: *natural_key* 36/0/0, natural_keys 17/0/0, compliance_suite_inventory 2/0/0, conjoined_event_tenancy_compliance 20 passed and 2 skipped, 0 failed. The session-switching fact two_tenants_starting_one_stream_id_in_a_single_save_get_two_streams passes. The 2 skips are the default-tenant-disabled facts, which were skipped before this PR.

🤖 Generated with Claude Code