Harden checker scheduling, state ordering, workflows, and dashboard

[ivanvyd]

An API request that rescheduled a checker could cancel after success and permanently stop that checker's timer. Two replicas sharing a checker name could also let an earlier-started check finish late and overwrite a newer result. This PR keeps timer lifetime with the scheduler and host, and prevents stale results from rewinding persisted health state.

The same change set addresses the related audit findings:

• Preserve the existing public timer constructor for compiled consumers and stop local timers when the host stops.
• Avoid individual state-store reads for missing built-in checker states after a bulk read; AI diagnosis looks up only the named checker, reuses its state history for built-in checkers, and preserves custom checker history behavior.
• Pass release and lock-refresh workflow inputs to Bash through environment variables instead of interpolating them into shell source.
• Raise faint and paused dashboard text contrast to at least 4.5:1 on the relevant opaque surfaces, and label checker status, rate, pulse history, and last-run values for assistive technology.

The timer, state-ordering, read-amplification, constructor, telemetry, and dashboard regressions were observed before their fixes. A Release solution build passed with zero warnings. Direct xUnit runners on both .NET 8 and .NET 10 passed 637 unit tests each (590 passed, 47 Docker-backed tests skipped); the browser E2E suite passed 54/54. A local dashboard sweep at 375, 768, 1280, and 1920 px, plus throttled mobile, showed no overflow or console errors, and keyboard activation of a checker worked. Workflow YAML parsed successfully. The machine has SDK 10.0.400 while global.json requires 10.0.401, so the standard dotnet test command could not run; the built xUnit executables were used instead.

This PR does not publish a package or deploy production. The separate landing hardening is in ivanvyd/landings#22; its visible version remains 4.1.3, the latest published release.