feat(studio-cp,oab-mcp): k8s roster observe + runtime_context dispatch

[brettchien]

Summary

studio#146 slice 2 of 4 (k8s fleet roster/observe UI). Builds on slice 1 (#148, merged) to add the actual k8s observe path, and wires oab-mcp to dispatch to it for a k8s-runtime fleet.

crates/studio-cp — the k8s counterpart to the existing ECS read-model (observe_services/observe_deployment/build_deployment):

• observe_k8s_services(context, namespace): lists OAB-managed Deployments (oab/name label present) in a namespace — the k8s counterpart to observe_services.
• observe_k8s_deployment(context, namespace, service): one Deployment's replica counters + per-Pod phase, listing its Pods by the Deployment's own selector and mapping each through agent_lifecycle::k8s::K8sDriver (from feat(agent-lifecycle): k8s RuntimeDriver — projects Pod state onto the 6-state model #148) — same Deployment/InstancePhase output type ECS produces, no new console-facing shape.
• Pod → phase mapping reads phase, the Ready condition, whether a readiness probe is declared, and CrashLoopBackOff from container statuses — the native signals K8sDriver::project expects.

crates/oab-mcp — deploy_list/deploy_get/runtime_context now check the named fleet's runtime and branch to the k8s path (the new studio-cp functions, and the already-existing observe_k8s_identity for runtime_context) before falling through to the existing ECS logic. target() itself is untouched and still rejects k8s fleets — every other tool (scale/delete/apply/provision/events) has no k8s dispatch yet and should keep erroring loudly rather than silently acting against an empty cluster string. Tool descriptions updated to mention k8s support where applicable.

A naming wrinkle worth flagging explicitly: a k8s Deployment's own on-cluster resource name is oab-{slug(name)} (k8s_driver.rs's k8s_deployment_name, no namespace embedded — namespace is a first-class k8s field there). But fleets.toml's members array stores oab-{namespace}-{name} (console/src/deploy.ts's service = oab-${namespace}-${name}, written unconditionally for both runtimes). observe_k8s_services/observe_k8s_deployment reconstruct the fleets.toml form from the Deployment's oab/name label (which k8s_driver.rs::build_deployment sets to the original, un-slugified agent name) rather than using the on-cluster resource name, so FleetBinding::includes matches the same way it does for ECS's full-vs-short dual match. Flagging this because it's exactly the kind of naming mismatch that reads as "roster shows nothing" if missed.

runtime_context's k8s response shape: cluster is null, context/namespace are set instead of it — noting this for whoever picks up slice 4 (console read-path wiring), since that's the exact JSON shape it'll need to branch on.

Test plan

• cargo test -p studio-cp --lib — 35/35 passing (no regressions; the new k8s observe functions need a live cluster so aren't unit-tested here, same as observe_k8s_identity/list_namespaces before them)
• cargo test -p oab-mcp --lib — 3/3 passing (incl. catalog_advertises_the_named_tools, unaffected by the description-only tool-schema edits)
• cargo check -p studio-cp -p oab-mcp — clean
• Full workspace build/test — left to CI per this sandbox's resource constraints (aws-sdk-ec2 OOM'd the sandbox's default test-profile debuginfo; worked around locally with CARGO_PROFILE_TEST_DEBUG=0, not something to carry into the PR itself)
• Manual: exercise deploy_list/deploy_get/runtime_context against a real k8s-runtime fleet — no live cluster available in this sandbox

🤖 Generated with Claude Code