perf(api): /v1/gpu-breakdown clinic flagged memory as 'data' category (potential leak)

Problem

GET /v1/gpu-breakdown was flagged by clinic doctor with the data category, which indicates a potential memory leak or unusual memory allocation pattern. Clinic specifically flagged heapTotal and heapUsed as data issues.

Clinic Doctor Results

• Event loop delay: avg 0.95ms (low)
• CPU: avg 0.02% (very low — endpoint barely processes)
• Throughput: 0 req/s effective (only 20 requests completed in 10s)
• RSS: 335MB (low)
• Clinic category: data — indicates abnormal memory behavior
• Clinic flags: cpu=performance, memory.heapTotal=data, memory.heapUsed=data

Analysis

The data category from clinic is unusual and suggests the endpoint's memory usage pattern doesn't follow normal request/response cycles. The very low throughput (20 requests in 10s) combined with data memory flags could mean:

• Memory is allocated but not freed between requests
• The endpoint triggers a large one-time computation that isn't properly garbage collected
• Cache warming creates a large allocation spike

Suggested investigation

• Review the GPU breakdown service for memory allocation patterns
• Check if large intermediate data structures are created and retained
• Profile with --heap-prof to identify allocation sources

Files

• apps/api/src/gpu/ (routes + services for gpu-breakdown)

Context

Part of the API event loop performance audit.

Start by reviewing the routes and services under apps/api/src/gpu/ for the /v1/gpu-breakdown endpoint. Run the suggested Clinic profiling with --heap-prof and inspect heapTotal and heapUsed allocation sources. Done means identifying and addressing the abnormal memory behavior, then confirming the endpoint no longer produces the flagged pattern.