refactor(discovery): add NodeDiscovery interface and consolidate node listing

[ev-shindin]

Summary

Pure refactor of internal/discovery/k8s_with_gpu_operator.go — no behavior change for existing callers. Extracts a single internal node-listing helper, adds a new NodeDiscovery interface that exposes per-node labels, and re-implements the existing Discover and discoverNodeGPUTypes methods as projections over the new helper.

This unblocks upcoming label-aware features (namespace-scoped limiter, node-aware bin packing) that need node.Labels alongside accelerator capacity, without forcing a third near-identical vendor-loop node query.

Motivation

K8sWithGpuOperator currently has two near-identical vendor-loop node queries:

• Discover() (lines 33-99) — builds map[nodeName]map[model]AcceleratorModelInfo
• discoverNodeGPUTypes() (lines 147-188) — builds map[nodeName]string

Both loop over vendors, build the same selector, list nodes, iterate results, and differ only in the projection. Upcoming work (namespace-scoped limiter, epic limiter improvements) needs a third projection — per-node info including node.Labels. Adding a third independent vendor-loop would lock in the duplication pattern. This PR consolidates node discovery into a single primitive with multiple public projections.

Behavior preservation

This is a pure refactor. Verified invariants:

• Discover() output shape and contents unchanged — existing tests pass without modification
• DiscoverUsage() unchanged — still uses discoverNodeGPUTypes internally
• discoverNodeGPUTypes() multi-vendor tie-break preserved (intel > amd > nvidia if multiple labels present). Locked down by new TestDiscoverNodeGPUTypes_MultiVendorNode_LastWins.
• WVA_NODE_SELECTOR handling unchanged; error on invalid selector is now explicitly tested
• Vendor iteration order (nvidia.com, amd.com, intel.com) preserved
• Multi-vendor nodes merged into a single NodeInfo entry with both accelerators
• Nodes without Allocatable resource still included with Count=0
• Non-GPU nodes excluded from DiscoverNodes output (same as Discover)

Tests

All existing discovery tests pass unchanged (behavior-equivalence check).

Eight new tests added:

• TestDiscoverNodes_SingleVendor — basic labels + accelerators capture
• TestDiscoverNodes_MultiVendorNode — merged single entry
• TestDiscoverNodes_RespectsWVANodeSelector — sharding env var respected
• TestDiscoverNodes_NodeWithGPULabelButNoAllocatable — Count=0 preserved
• TestDiscoverNodes_EmptyCluster — empty input → empty output
• TestDiscoverNodes_ExcludesCPUOnlyNodes — filters consistent with Discover
• TestDiscoverNodes_InvalidWVANodeSelectorReturnsError — error path tested for both DiscoverNodes and Discover
• TestDiscoverNodes_LabelsAreIndependentCopy — mutation-safety contract
• TestDiscoverNodeGPUTypes_MultiVendorNode_LastWins — behavior-preservation regression test

Verification

• go build ./... — clean
• go test ./internal/... — all packages pass (including envtest suites for actuator, controller, engines/saturation)
• golangci-lint run ./internal/discovery/... ./internal/engines/pipeline/... — 0 issues

Non-goals

• No new behavior for existing callers
• DiscoverNodes not consumed anywhere yet; consumption happens in the follow-up namespace-scoped limiter PR
• No changes outside internal/discovery/ and the pipeline test mock

Follow-ups (out of scope)

• Namespace-scoped GPU inventory and limiter (issue) — will consume DiscoverNodes
• Epic: Limiter improvements

Test plan

• go build ./... clean
• go test ./internal/discovery/... — 18/18 pass
• go test ./internal/engines/pipeline/... — all pass (mock updated)
• go test ./internal/engines/saturation/... (envtest) — all pass
• golangci-lint run ./internal/discovery/... ./internal/engines/pipeline/... — 0 issues
• CI pipelines green on PR

[ev-shindin]

/ok-to-test

[github-actions]

🚀 Kind E2E (full) triggered by /ok-to-test

View the Kind E2E workflow run

[github-actions]

🚀 OpenShift E2E — approve and run (/ok-to-test)

View the OpenShift E2E workflow run

[github-actions]

GPU Pre-flight Check ✅

GPUs are available for e2e-openshift tests. Proceeding with deployment.

Resource	Total	Allocated	Available
GPUs	50	45	5

Cluster	Value
Nodes	16 (7 with GPUs)
Total CPU	993 cores
Total Memory	10383 Gi
GPUs required	4 (min) / 6 (recommended)