fix(benchmark): wire measured profiles into recommender

[Mog9]

Problem

The recommender was using 100% estimated data — 28 benchmark profiles, all marked "source": "estimated". Meanwhile, 4 measured profiles (real benchmark data from actual RTX 4070 runs) were sitting in benchmarks/profiles/measured/ but never loaded into the system.

Three checkpoints were blocking the measured data:

1. BenchmarkStore used glob("*.json") which only finds files directly in benchmarks/profiles/, not in subdirectories like measured/
2. GpuTier enum had no RTX_4070 entry, so even if loaded, the profiles would fail validation
3. RuntimeRegistry didn't list rtx4070 as a supported GPU tier for vLLM, so the candidate generator would filter them out

Changes

app/services/benchmark_store/store.py

Changed glob("*.json") to rglob("*.json") so the store recursively scans subdirectories and finds the measured profiles.

app/schemas/hardware.py

Added RTX_4070 = "rtx4070" to the GpuTier enum and added a GpuSpec entry to GPU_CATALOG:

• VRAM: 12 GB
• FP16 TFLOPS: 44
• Memory bandwidth: 432 GB/s
• Hourly cost: $0.00 (local GPU, matching the measured profiles)

app/services/runtime_registry/registry.py

Added "rtx4070" to vLLM's supported_gpu_tiers list. Not added to TensorRT-LLM (no TRT-LLM data for RTX 4070).

tests/test_engine.py

Updated 2 tests that broke because RTX 4070 at $0/hr is now a valid candidate:

• test_cost_optimized_recommendation — now accepts "rtx4070" as a valid budget GPU
• test_recommendation_response_shape — changed > 0 to >= 0 for hourly cost (local GPU is free)

Benchmarks Now Active

The 4 measured profiles are now flowing through the recommender:

Model	GPU	Backend	Quantization	Tokens/sec	Source
qwen2.5-7b	RTX 4070	vLLM	AWQ	79.4	measured
qwen2.5-3b	RTX 4070	vLLM	FP16	66.4	measured
llama3.1-8b	RTX 4070	vLLM	AWQ	81.0	measured
llama3.2-3b	RTX 4070	vLLM	AWQ	113.0	measured

Total profiles: 32 (was 28)
Measured profiles: 4 (was 0)
Estimated profiles: 28 (unchanged)

Test Results

149 passed, 1 skipped

All existing tests pass with the updated assertions.

Impact

• The recommender now has real benchmark data for RTX 4070, not just estimates
• Users can get recommendations for local GPU deployments (RTX 4070 at $0/hr)
• Cost-optimized workloads may now pick RTX 4070 over cloud GPUs
• No breaking changes to existing profiles or scoring logic