Live compression tests have two ~200-second midnight blind windows from now()-relative chunk seeding #1972

Three live compression tests seed chunks with now()-relative timestamps, which makes chunk PLACEMENT depend on the time of day the suite runs. There are two ~200-second windows around midnight UTC where the tests fail deterministically - not flaky in the reused-database sense, just blind windows the clock walks through every day.

Evidence (caught live tonight)

• dev run 30673966983 (head 3695dcf): green, darling-pg started 23:50 UTC 2026-07-31.
• PR Export the viewer's whole config folder instead of documenting it (#1953) #1968 run 30674344636 (head a1cc3d2, byte-identical TimescaleSupportTests.cs): failed at 00:01:38 UTC 2026-08-01.

EndToEnd_CompressionRun_TakesEligibleChunksAndLeavesYoungOnes_AgainstDevPostgres - Assert.Equal() expected 1, actual 2, at TimescaleSupportTests.cs:740 (the UNcompressed chunk count). A rerun outside the window passed with no code change.

Mechanism

The tick tables are day-aligned 1-day chunks. The young-chunk seed is:

INSERT ... SELECT now()::timestamp - (g || ' seconds')::interval ... generate_series(1, 200)

which writes rows spanning [now-200s, now-1s]. Run between 00:00:00 and 00:03:20 UTC, that span straddles the day boundary, so the "one young chunk" is actually TWO chunks (yesterday's tail + today's). Total chunks = 3, uncompressed = 2, and line 740 fails exactly as observed.

The old-chunk seeds have the symmetric problem in the other direction:

INSERT ... SELECT now()::timestamp - INTERVAL 'N days' + (g || ' seconds')::interval ...

spans [now-Nd+1s, now-Nd+200s]. Run between 23:56:40 and 23:59:59 UTC, the +200s end crosses midnight N days back, splitting the old data into two chunks - both closed, both past the eligibility delay, so the policy compresses BOTH and the compressed-count assert (line 739) fails instead.

Affected sites (all in Darling/Darling.Tests/TimescaleSupportTests.cs)

Nightly scheduled darling-pg runs and evening PR pushes can both land in these windows; tonight's did.

Fix direction

Anchor the seeds to midday instead of now():

date_trunc('day', now()) - INTERVAL 'N days' + INTERVAL '12 hours' + (g || ' seconds')::interval

A +/-200s span around midday can never cross a day boundary, chunk placement becomes deterministic for any run time, and no assertion changes are needed. The young chunk keeps its "still open, still young" property (its chunk is today's) - worth one careful look at the range_end > now() - CompressAfterDays assert at line 749 while touching it, but midday-today satisfies it with hours to spare.

Mechanical fix, four lines plus comments; the only care point is keeping the young/old semantics each test pins.