fix(memory): allow migration-aware runtime health startup

[Memtensor-AI]

Summary

• Extend the default Memory runtime activation health-check budget from 15 seconds to 120 seconds.
• Keep the timeout configurable through the Memory CLI.
• Pass the existing 120-second budget from Desktop startup.
• Cap retry backoff by the remaining health-check deadline.
• Preserve complete cleanup when first installation or activation truly fails.

Why

Memory 2.1.0 may migrate a large database from v6 to v7 during startup. For a large user database this can take about 20–22 seconds, so the previous 15-second installer deadline stopped the service too early and rolled back a successful installation.

Behavior

• Health checks are not skipped; activation still requires a healthy, protocol-compatible runtime.
• A genuinely expired health-check budget still rejects and rolls back the installation.
• Existing-runtime upgrades and first-install flows both retain their cleanup behavior.
• Retry sleeps are capped to the remaining healthCheckTimeoutMs budget and never oversleep the deadline.
• CLI callers can configure the timeout, and Desktop passes its existing 120-second budget.

Validation

• Project Harness Full/T3: 12/12 passed.
• Run: run-1788491788350-85111f1b-ff3f-4346-8bbf-4aab3bd578cc
• Base: release/v1.1.2 at d7096887a0dcfd6748f23b5c1ba3b5a0e0488d57
• Coverage includes delayed activation, timeout rollback, first-install cleanup, CLI/Desktop propagation, and deadline-capped backoff.
• Follows merged PR test(windows): track packaged runtime manifest helper #353.

[Memtensor-AI]

Project Harness review handoff marker. Only reviews submitted strictly after this GitHub timestamp qualify.