Make the PgTarget anomaly and blocking tests independent of the wall clock (#4274)

[erikdarlingdata]

Closes #4274.

Why

Both tests derived their analysis window from a raw DateTime.UtcNow minute
(TruncateToMinutes(DateTime.UtcNow).AddMinutes(-1)). AnomalyGate.EvaluateTiles
(PerformanceMonitor.Analysis/AnomalyGate.cs) skips any target-local-hour tile under
AnomalyThresholds.MinTileSamples (3) before it is even scored (design: #3653 A8 option B). Depending on
the wall clock at test-run time, the tile holding the seeded spike (anomaly test) or the window's sole
zero-blocked minute (blocking test) could land at 1-2 samples, get excluded, and let a different tile "win"
worst-scoring — misreporting the peak, or dragging the mean below a pinned assertion range.

Diagnosis, confirmed with numbers

Added a test-only clock seam (DARLING_TEST_NOW_UTC, read by a new SimulatedUtcNow() helper in each
file; unset in every normal run, CI included) so the exact CI failures could be reproduced on demand
against a real dev Postgres, instead of waiting for the wall clock to land on a bad minute.

Anomaly test, TheAuroraWaitProfile_..._AgainstDevPostgres, old formula, anchor minute :02 (spike's
own tile sized 2, under the floor of 3):

Assert.Equal() Failure: Values are not within 3 decimal places
Expected: 3200 (rounded from 3200)
Actual:   1500 (rounded from 1500)

Exact match to the two CI failures in #4274. At anchor minute :31 (tile sized 31, safely over the floor)
the same old formula passed.

Blocking test, ThirtyOneDaysOfLightBlocking_..._AndTheAnomalyInOneIncident, old formula, anchor
minute :49 (windowStart's own tile — the one holding the window's sole zero-blocked minute — sized
60 - 49 = 11):

Assert.InRange() Failure: Value not in range
Range:  (2.7999999999999998 - 3)
Actual: 2.7272727272727271

Exact match to the CI failure in #4274, digit for digit. At anchor minute :21 (tile sized 39) the same old
formula passed.

The under-report, not acted on: at the anomaly test's bad anchor, before reverting to the fix, I
probed the fact's other metadata. current_ms_per_sec read 1500 (wrong), but window_peak read 3200
(right) and window_samples (of the winning, wrong tile) read 60 — WindowTiles.WholeWindow sums the
unfiltered tile list, so the true peak survives in a less prominent field even when the headline number
doesn't. This is real product behavior under #3653 A8 option B's design (a too-small edge tile never gets
scored, full stop), not a test bug, so AnomalyGate and MinTileSamples are untouched here. Posted the
numbers as a question for the coordinator on #4274 rather than changing product code.

What changes

Both files get the same three additions, plus one anchor-line swap and one refreshed comment:

• TruncateToHour(DateTime value) — truncates to the top of the hour (mirrors the existing
  TruncateToMinutes already in both files).
• SimulatedUtcNow() — returns DateTime.UtcNow, or an ISO-8601 parse of DARLING_TEST_NOW_UTC when
  that env var is set. The test-only clock seam the brief asked for.
• AnchorEndUtc() — TruncateToHour(SimulatedUtcNow()).AddMinutes(-1). Always :59 of the PRIOR hour,
  regardless of what minute SimulatedUtcNow() returns, so the tile holding the seed is never undersized.
• The two TruncateToMinutes(DateTime.UtcNow).AddMinutes(-1) anchor lines now call AnchorEndUtc().
• The blocking test's stale comment ("was pinned 2.9-3.0; the worst tile's hour reads 2.8333...") is
  replaced with the current mechanism and the now-deterministic value (see the table below).
  Assert.InRange(anomaly.Metadata["avg_blocked_sessions"], 2.8, 3.0) itself is untouched — not widened,
  not tightened — because the range predates this fix and 3.0 sits inside it.

No product code changes. No CHANGELOG entry (test-only).

Two other tests in these same files (PgTargetBlockingTests.cs lines 913 and 1125, in a different method
than the one fixed here) share the same raw-minute TruncateToMinutes(DateTime.UtcNow).AddMinutes(-1))
anchor pattern but were not named in #4274 and were not observed failing — left alone, in scope for a
follow-up if they ever flake.

Test plan

Rig: Postgres 55972 in C:\GitHub\worktrees\rig-fx2 (fresh extract of pg-runtime.zip), timezone/
log_timezone pinned to UTC, timescaledb.max_background_workers/max_worker_processes at CI's 74/85.

Proof matrix, both tests, final (fixed) build, AnchorEndUtc() driven by DARLING_TEST_NOW_UTC:

Anchor (DARLING_TEST_NOW_UTC)	windowEnd after AnchorEndUtc()	Anomaly test	Blocking test
2026-09-25T10:00:30Z (minute :00)	09:59:00	pass	pass
2026-09-25T10:01:15Z (minute :01)	09:59:00	pass	pass
2026-09-25T10:15:00Z (minute :15)	09:59:00	pass	pass
2026-09-25T10:30:00Z (minute :30)	09:59:00	pass	pass
2026-09-25T10:45:00Z (minute :45)	09:59:00	pass	pass
2026-09-25T10:59:59Z (minute :59)	09:59:00	pass	pass
2026-09-25T02:15:00Z (crosses UTC midnight: window runs 2026-09-24 21:59 to 2026-09-25 01:59)	2026-09-25T01:59:00	pass	pass

(An eighth anchor, 2026-09-26T02:15:00Z, was tried first for the midnight case and produced an unrelated
"invalid" analyze_server status — that anchor was in the future relative to the real clock, and the
service correctly rejects an as_of in the future. Not a tile bug; swapped for the past-dated midnight
anchor above.)

Probed the fixed formula's avg_blocked_sessions directly (temporary forced-mismatch assertion, reverted)
at two of the anchors above: both read exactly 3.0. windowStart's own tile (the one that used to drag
the mean down) is now always sized 1 — under the floor of 3 — so it never scores, and every tile that does
score is a clean 3.0.

• dotnet build Darling/Darling.Tests/Darling.Tests.csproj -c Debug — Build succeeded, 0 Warning(s), 0 Error(s).
• git merge origin/dev — clean, no conflicts.
• Full Darling.Tests.exe, fresh darlingtest, real clock (no DARLING_TEST_NOW_UTC), foreground, once: Total: 14070, Errors: 0, Failed: 0, Skipped: 49, Not Run: 1, Time: 731.212s. The 49 skips and 1 not-run are pre-existing (gated live tests needing other env vars / runtimes this rig didn't set up) and unrelated to the two files this PR touches.
• Rig stopped (pg_ctl stop) after the run.

Coordinator check

• The CI flake: PgTarget anomaly/blocking worst-tile assertions fail on first attempts unrelated to the change #4274 comment above flags the current_ms_per_sec/window_peak under-report as a question, not a
  ruling — please confirm whether it needs a follow-up issue or is accepted design.
• Confirm the two same-pattern tests at PgTargetBlockingTests.cs:913 and :1125 (out of scope here) are
  worth a follow-up if they start flaking too.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TszxYhJJbTEh4LrZ56NYo3