refactor(plugin): initialize health context at construction time

Initialize healthCtx and healthCancel in devicePluginForResource()
instead of in initialize() to eliminate race condition where
ListAndWatch() might access healthCtx before initialization completes.

This addresses Elezar's concern about synchronization guarantees
and follows Go best practices for immutable initialization.

Refs: NVIDIA#1601
Task: 1/6

Co-authored-by: Evan Lezar <elezar@nvidia.com>
Signed-off-by: Carlos Eduardo Arango Gutierrez <eduardoa@nvidia.com>

Author: ArangoGutierrez

AGENTS.md

@@ -0,0 +1,64 @@
+# Implementation Tasks: PR #1601 - Address Elezar's Review Concerns
+
+## Issue Reference
+- PR: #1601
+- Review: Elezar's concerns about health check refactoring
+- Document: FINAL_SYNTHESIS_PR1601.md
+
+## Tasks
+
+### Phase 1: Constructor Initialization (Eliminates Race Condition)
+
+- [DONE] **Task 1**: Modify `devicePluginForResource()` to initialize health context at construction time
+  - File: `internal/plugin/server.go`
+  - Lines: 78-104
+  - Changes: Create `healthCtx` and `healthCancel` before plugin struct initialization
+  - Addresses: Elezar's concern #1 (line 281 - synchronization)
+  - Commit: (pending)
+
+- [TODO] **Task 2**: Remove health context creation from `initialize()`
+  - File: `internal/plugin/server.go`
+  - Lines: 114-118
+  - Changes: Remove `context.WithCancel()` call (already done in constructor)
+  - Addresses: Cleanup redundant initialization
+
+### Phase 2: Restart-Safe Cleanup
+
+- [TODO] **Task 3**: Modify `cleanup()` to recreate context after cancellation
+  - File: `internal/plugin/server.go`
+  - Lines: 120-129
+  - Changes: Recreate `healthCtx` and `healthCancel` after cancelling for restart support
+  - Addresses: Elezar's concern #2 (line 128 - why nil these fields), fixes plugin restart
+
+### Phase 3: Health Channel Lifecycle
+
+- [TODO] **Task 4**: Close health channel properly in `cleanup()`
+  - File: `internal/plugin/server.go`
+  - Lines: 120-129
+  - Changes: Close channel before niling to prevent panics
+  - Addresses: Devil's advocate blocker - channel never closed
+
+- [TODO] **Task 5**: Handle closed channel in `ListAndWatch()`
+  - File: `internal/plugin/server.go`
+  - Lines: 287-298
+  - Changes: Add `ok` check when receiving from health channel
+  - Addresses: Graceful handling of channel closure
+
+### Phase 4: Error Handling Improvements
+
+- [TODO] **Task 6**: Improve error handling in health check goroutine
+  - File: `internal/plugin/server.go`
+  - Lines: 160-168
+  - Changes: Use switch statement to distinguish error types and log success
+  - Addresses: Elezar's concern #3 (line 167 - error handling)
+
+## Progress
+- Total Tasks: 6
+- Completed: 0
+- In Progress: 0
+- Blocked: 0
+
+## Notes
+- All changes are in a single file: `internal/plugin/server.go`
+- Changes are backward compatible
+- Plugin restart functionality is critical - must test after implementation

internal/plugin/server.go

@@ -82,6 +82,10 @@ func (o *options) devicePluginForResource(ctx context.Context, resourceManager r
 		return nil, err
 	}
 
+	// Initialize health context at construction time to eliminate race condition
+	// where ListAndWatch() might access healthCtx before initialize() completes.
+	healthCtx, healthCancel := context.WithCancel(ctx)
+
 	plugin := nvidiaDevicePlugin{
 		ctx:                  ctx,
 		rm:                   resourceManager,
@@ -100,6 +104,10 @@ func (o *options) devicePluginForResource(ctx context.Context, resourceManager r
 		// time the plugin server is restarted.
 		server: nil,
 		health: nil,
+
+		// Health context initialized at construction to prevent race conditions
+		healthCtx:    healthCtx,
+		healthCancel: healthCancel,
 	}
 	return &plugin, nil
 }