Merge train 4843-18 (CI only; not for merge)

[erikdarlingdata]

Merge train, not for merge. CI runs once on dev plus the squash of each PR below, in this order. When it is green, each is squash-merged by itself in the same order, dev's tree is checked against this branch's, and this PR is closed unmerged.

Part of #4966, #5097, #5098 and #5208. Carries migration rung V163.

PRs in this train, in order

• Tests: the system-health and default-trace window-notice tests get their own live-postgres classes (part of Say where the data starts on every time-ranged tab, in Lite and both Darling viewers #4966) (Tests: the system-health and default-trace window-notice tests get their own live-postgres classes (part of #4966) #5214)
• DARLING: the store keeps an hourly history of its own pg_stat_statements deltas (V163, part of Slow reads and slow collector runs can't be diagnosed from the product: record each slow call's arguments, fix the stall probe's waits, keep store statement history, and export a diagnostics bundle #5097) (DARLING: the store keeps an hourly history of its own pg_stat_statements deltas (V163, part of #5097) #5203)
• CI: per-test Lite timings as an artifact, and temporary cap guards while Lite.Tests on hosted Windows runners grew from ~9½ to ~45 minutes in five days; the unsharded nightly run now exceeds its 90-minute cap #5208 is open (CI: per-test Lite timings as an artifact, and temporary cap guards while #5208 is open #5211)
• Web: the event grids' data-start note uses the earliest event shown (part of Say where the data starts on every time-ranged tab, in Lite and both Darling viewers #4966) (Web: the event grids' data-start note uses the earliest event shown (part of #4966) #5197)
• MCP: Darling's blocking stats and PostgreSQL deadlock and log event tools say where their data starts (part of Say where the data starts on every time-ranged tab, in Lite and both Darling viewers #4966) (MCP: Darling's blocking stats and PostgreSQL deadlock and log event tools say where their data starts (part of #4966) #5205)

Train head 0b92d8248, tree 7f7f2b1e5, base dev 1acb3d710.

The first member goes first on purpose, ahead of the rung. It fixes a live-store hygiene guard that dev itself trips: two window-notice test files were added as partial class parts of existing live classes, while the serialising attribute sits only on the other part, so the guard's per-file scan flags them. Until that is fixed, every train carries a standing red, and a standing red is how a real failure gets waved through. A Windows VM confirmed the guard RUNS and passes on that fix (11 tests, 0 failed) and runs and fails on plain dev (11 tests, 1 failed), and that the offender list is identical on dev and on dev plus another member here, so the fix covers the whole set rather than tonight's symptom.

The rung member is second; nothing in this train depends on its schema, so that ordering is free.

Local merged-tree checks: 1,800 tests, 0 build warnings. Two classes fail, both matching dev at this commit: the store-upgrade class at 9, and the hygiene class, which on macOS cannot load a Windows presentation assembly and so cannot speak to the fix at all — that is what the VM run above is for. The repository-source scan over 400 classes reports no new failure.

[erikdarlingdata]

Superseded by train 18b. One member introduces a new offender to the live-store hygiene guard that this train's first member fixes, and is the prime suspect for a second, unrelated live-test failure that dev does not have. The other four members ride 18b. Closing unmerged.