[aranadive]

Summary

Add optional, read-only HTTP health probes to the standalone KVCR service so deployments can observe service and Guard progress without claiming pools or changing leases.

Stacked on #116: this draft targets pull-request/116, the upstream copy of PR 116 at 95b3cb2. The review diff contains only the health-probe commit (614b1f8). After #116 merges, rebase as needed and retarget this PR to main; do not merge into the temporary base branch.

Companion deployment draft: ai-dynamo/dynamo#17290.

Changes

• Enable GET /live and /ready with --health-port; bind to loopback by default. Both return 503 on known failures, shutdown, or stalled Unix accept/Guard actor loops.
• Support an optional HTTP companion dependency and configurable progress timeout; validate options before pool allocation and stop the listener during shutdown.
• Keep background preparation nonblocking for readiness. A promotion waiting for preparation can report stalled progress independently of the Guard startup budget.
• Document probe limits and coordinated recovery: readiness does not establish inference readiness, and restarting only the Guard while engines map its pools is unsafe.

Validation

• ruff check src tests and ruff format --check src tests: passed.
• Full tests/unit suite against the rebased source in the qualified Linux/NIXL container: 649 passed. A separate host-native run had four failures, including NIXL backend/FlexIO errors; the container run is the passing reference.
• Added coverage for invalid options, HTTP routes, dependency failure, shutdown, read-only lease observation, stalled loops, and preparation/promotion health transitions.
• Every tracked file matches the previously GPU-tested combined PR 116/health snapshot. Earlier integrated testing with a compatible vLLM overlay covered GPU KV reuse and Guard recovery at 2 GiB and 1 TiB pool capacities; 1 TiB is capacity, not transferred payload. GPU tests were not rerun for the rebase.