CI: run the store-upgrade tests on pull requests that touch the store's schema or upgrade path

[erikdarlingdata]

Why

Two gated end-to-end tests exist only to catch a store-upgrade regression before it ships: an
upgraded-in-place major upgrade over a real store with a hypertable, TOAST-sized plan XML, a
continuous aggregate and a compressed chunk, checksum-compared before and after; and a same-major
runtime swap with no pg_upgrade involved. Both are gated on three environment variables
(DARLING_TEST_PGRUNTIME_OLD/_NEWZIP/_PREVIOUS) that point at a fixture nightly CI builds and
sets every night — and only there. Pull-request CI never built that fixture and never set those
variables, so both tests silently skipped on every single pull request, including ones that changed
the exact code they exist to cover.

What changes

• A new store-upgrade: path filter in .github/darling-paths-filter.yml, covering the store
  schema, the service project holding the upgrade orchestration and the managed-conf migration
  machinery it drives, the two fixture-build scripts, and the workflow/filter files themselves (the
  same self-validating reasoning the existing darling: filter already uses for build.yml).
• build.yml's PG test job now builds the same upgrade fixture nightly builds, and sets the same
  three environment variables on the test step, on every shard, so a class's shard doesn't decide
  whether it runs — only when a pull request's changed files match the new filter. Every PR that
  doesn't touch this filter's paths is unaffected; no new step runs, so the build time for those legs
  is unchanged. nightly.yml is untouched.
• A pin in DarlingPathFilterGateTests asserting the filter exists and covers the store schema
  directory and both fixture scripts; a companion test asserts an injected removal of the schema entry
  is detected. Both run green on this branch; the missing-entry case was verified RED by hand before
  restoring the file (see below).

Test plan

• dotnet build Darling/Darling.Tests/Darling.Tests.csproj -c Release -p:EnableWindowsTargeting=true:
  0 warnings, 0 errors.
• python3 -c "import yaml; yaml.safe_load(...)" on both workflows and the filter file: clean.
  actionlint on build.yml: no new findings (the one pre-existing shellcheck note is in
  nightly.yml, which this PR does not touch).
• DarlingPathFilterGateTests run in-process on the built assembly: 8/8 green, including the two new
  facts.
• Mutation, run and reverted locally: removed the Storage entry from the new filter's pattern
  list, rebuilt nothing (the test reads the YAML at run time), reran the same class — both new facts
  failed as expected (TheStoreUpgradeGate_CoversTheSchemaAndTheFixtureScriptsAndTheServiceProject and
  TheStoreUpgradeGate_MissingTheStorageEntry_IsDetected), then restored the file and reran — 8/8 green
  again.
• Proven in pull-request CI with three throwaway draft PRs (Not for merge: CI proof (upgrade-touch-storage) #4497, Not for merge: CI proof (upgrade-touch-lite) #4498, Not for merge: CI proof (upgrade-mutation) #4499, since closed and their branches deleted), each one commit on top of this branch:
  • Not for merge: CI proof (upgrade-touch-storage) #4497, a comment-only change to PgMigrations.cs: every PG shard ran the fixture steps (the cached download ~6–10 s, the previous-major runtime build ~26 s) and passed.
    • The shards' [SKIP] lists hold none of the upgrade tests. Shard 0 skipped 6 (the live-SQL-Server and live-PostgreSQL-target E2Es, and a cert shape), shard 1 skipped 2 (the PG-log live tests), and shard 2 skipped 0.
    • So DarlingStoreUpgradeTests, DarlingManagedPostgresTests and ManagedConfUpgradePathTests ran their gated facts.
    • Totals 5,088 + 4,947 + 5,667, 0 failed.
  • Not for merge: CI proof (upgrade-mutation) #4499, a one-line break at the in-place upgrade's pg_upgrade call site (the serverOptions argument dropped; the argument-builder unit pins don't reach that call):
    • it went RED in PR CI on exactly the gated upgrade tests: DarlingStoreUpgradeTests.UpgradeInPlace_OldMajorStoreWithRealData_UpgradesAndKeepsEverything_Gated (shard 2) and ManagedConfUpgradePathTests.UpgradeInPlace_OperatorLineBelowInclude_CarriedAndApplied_NoManagedKeyDuplicated_Gated (shard 1).
    • Everything else stayed green.
  • Not for merge: CI proof (upgrade-touch-lite) #4498, a comment-only Lite change, can't show the skip before this merges. A PR's diff against dev still contains this branch's workflow and filter files, which store-upgrade: includes on purpose (so a change to the gate exercises the gate). Not for merge: CI proof (upgrade-touch-lite) #4498 therefore ran the fixture too, as designed.
    • When store-upgrade doesn't match, the new steps' if: is false and the env expressions give an empty string, which every gated test treats as unset (IsNullOrWhiteSpace → skip).
    • The first Lite-only PR after this merges shows it.
• The cost when it matches: ~33 s per PG shard for the fixture, and the shards run in parallel.

CHANGELOG

None: CI-only change.