CI: run the self-hosted log collector's two end-to-end tests on pull requests

[erikdarlingdata]

Why

Darling/Darling.Tests/PgLogEventsPipelineTests.cs has two live end-to-end tests gated on
DARLING_TEST_PG_LOG_TARGET and DARLING_TEST_PG_UTC_LOG_TARGET. Neither variable was set in any
workflow, so both tests skipped in every CI run since they were written.

What changes

build.yml's darling-pg job gets a fifth throwaway PostgreSQL cluster (port 5545), alongside the
existing store (5541) and three log-format clusters (5542/5543/5544):

• logging_collector = on, log_destination = 'stderr', log_timezone = 'UTC'
• log_line_prefix = '%m %u@%d [%p] ' (the %u@%d shape the second test's reader.GetName(1) == "log_timezone" assertion needs)
• log_connections = on, log_disconnections = on, log_lock_waits = on
• deadlock_timeout = '100ms' (inherited from Start-LogFormatCluster's shared defaults)

One cluster serves both tests: both need logging_collector on and UTC timezone, and the first
test's extra preconditions (log_connections, log_lock_waits) don't conflict with the second's.

A workload step immediately after the cluster starts puts one real event of each family the first
test asserts on into that cluster's own log, via psql.exe from the job's own pg-runtime bin:

• an ordinary connect (SELECT 1) — produces the connection/disconnection LOG pair;
• SELECT 1/0 — a real ERROR the server writes itself;
• a lock wait: a background job holds LOCK TABLE inside a transaction and sleeps 3s, while a
  second psql requests the same table and blocks past deadlock_timeout, producing the
  log_lock_waits "still waiting for" / "acquired" pair;
• a connect to a nonexistent database — a FATAL, distinct from the ERROR above.

Every workload call checks its own exit code and throws on an unmet precondition (a psql failure
that should have succeeded, or a success where a failure was expected), so a silently-quiet log
fails the step instead of failing the test with no explanation.

DARLING_TEST_PG_LOG_TARGET and DARLING_TEST_PG_UTC_LOG_TARGET are set on "Run Darling PG tests"
(the connection-string shape of the existing 5542-5544 vars), both comment blocks are updated to
name the new cluster and both tests, and the "Stop PostgreSQL" / failure-artifact steps are
extended to cover the new data directory and log file.

nightly.yml starts no log-format clusters today (only the 5541 store), so there is nothing to
mirror there; left unchanged.

Test plan

• python3 -c "import yaml,sys; [yaml.safe_load(open(f)) for f in sys.argv[1:]]" .github/workflows/build.yml .github/workflows/nightly.yml — clean.
• actionlint .github/workflows/build.yml — clean.
• dotnet build Darling/Darling.Tests/Darling.Tests.csproj -c Release -p:EnableWindowsTargeting=true — 0 Warnings, 0 Errors.
• Darling.Tests.dll -class Darling.Tests.PgLogEventsPipelineTests against a two-container docker rig with the 5545 settings: Total: 61, Failed: 0 (pre-mutation) / Total: 61, Failed: 13 (mutated, including both new E2Es).
• Proven in PR CI:
  • This PR (0e99b5f3a): the 5545 workload step passes, and all three Darling PG tests shards pass (5,674 / 4,962 / 5,103 run, 0 failed). The shard that hosts PgLogEventsLivePostgresTests now skips 0; before this change it skipped exactly these two tests. So both run and pass.
  • A throwaway draft (Not for merge: CI proof (log-target mutation) #4505, now closed) with LOG dropped from PgLogEntryAssembler's label alternation: the workload step passes, and TheSelfHostedCollector_ReadsTheTargetsOwnLog_EndToEnd goes RED, along with ThePipeline_StoresReadsFiltersAndDedupes_AgainstDevPostgres and the parser pins.
  • AgainstAUtcTarget_… stays green under that mutation, because its line is an ERROR, not a LOG.
  • A second throwaway draft (Not for merge: CI proof (UTC log-target mutation) #4507, now closed) with DateTimeStyles.AdjustToUniversal dropped from PgLogEntryAssembler's timestamp parse, so parsed times come back as local rather than UTC: AgainstAUtcTarget_TheLogQueriesReadTheSettingWithTheTail_AndStoreWhatTheServerWrote and TheSelfHostedCollector_ReadsTheTargetsOwnLog_EndToEnd both go RED on their OccurredAtUtc.Kind == Utc assertion.

CHANGELOG

None: CI-only change.