Validate MiniCPM-o image to speech on WSL GPU

docs/design/omni_edge_e2e_roadmap.md

@@ -27,6 +27,12 @@ use reached 11.17 GB. Native Windows InternVLA direct-policy runs separately
 passed on CPU and RTX 5090 Laptop with synthetic inputs; AMD and mobile cells
 do not inherit these passes.
 
+The WSL RTX MiniCPM-o image-to-text+speech path now passes for one 448×448
+synthetic red-square image under an explicit 0.61 thinker GPU budget. The
+previous 0.58 budget failed the 2,048-token KV admission gate; the successful
+run still force-killed stage 0 during shutdown and reported a shared-memory
+cleanup warning. Broader image quality and clean lifecycle remain open.
+
 ## Architecture and prerequisites
 
 Keep Omni's PipelineConfig, StageRuntime, StageClient, orchestration and admission
@@ -82,7 +88,7 @@ capability and lifecycle implementation; a state-handle type alone is insufficie
 | M1 | S25 Spark | Device-local Omni-compatible controller; embedding, prefill, continuous decode, ring/full KV, head and sampler; resident state | Reference checks, at least 128 output tokens, 512-window/1024-bucket transitions, cancellation, resident memory and sustained profile |
 | M2 | Qwen3-TTS | Resolve desktop playback and shutdown findings and CPU compatibility; then mobile talker/predictor plus GPU vocoder | Complete PCM and tail, history/chunk correctness, interruption/recovery, quality checks, no post-startup underruns and RTF below 1 in the declared workload |
 | M3 | AMD stages | Connect compatible real encoders or other useful coarse stages through current Omni workers; start from the existing 890M vision evidence | Actual node/device placement, numerical and task quality, complete downstream output, shared-memory and handoff costs; reject unhelpful splits |
-| M4a | MiniCPM-o 4.5 | Three-stage desktop text+WAV produced coherent text and nonzero WAV on RTX 5090 Laptop under WSL and native Windows, and WSL CPU, using separate constrained plans. The CPU plan needed vLLM 0.28/0.29 processor compatibility and used swap; 20 serial CPU requests measured p50 21.86 s/p95 22.97 s. Native Windows needed longer startup limits; 20 serial requests measured p50 18.38 s/p95 18.73 s with host RAM/pagefile pressure. Add encoder and persistent-state reference before mobile artifacts | Speech intelligibility/alignment, loading-peak admission, concurrency, separate text, image and audio-understanding suites, then combined streaming and interruption |
+| M4a | MiniCPM-o 4.5 | Three-stage desktop text+WAV produced coherent text and nonzero WAV on RTX 5090 Laptop under WSL and native Windows, and WSL CPU, using separate constrained plans. One WSL RTX synthetic image also passed through vision encoding to correct text and nonzero WAV after increasing the thinker GPU budget from 0.58 to 0.61; shutdown cleanup remains open. The CPU plan needed vLLM 0.28/0.29 processor compatibility and used swap; 20 serial CPU requests measured p50 21.86 s/p95 22.97 s. Native Windows needed longer startup limits; 20 serial requests measured p50 18.38 s/p95 18.73 s with host RAM/pagefile pressure. Add other encoders and persistent-state reference before mobile artifacts | Speech intelligibility/alignment, loading-peak admission, concurrency, broad image and audio-understanding suites, shutdown lifecycle, then combined streaming and interruption |
 | M4b | InternVLA-A1 | Real weights, observation preprocessing, shared prefix state, iterative action head and versioned action metadata; strict-load synthetic policy forwards now pass on WSL and native Windows, each on CPU and RTX 5090 Laptop | Reference action agreement on real observations, units/order/timestamps, observation age, stale-input handling and application-specific deadline reporting |
 | Independent | Qwen3.8-27B | Public Omni pipeline binding; text, then image/video and longer contexts; same-model CPU/iGPU artifact and explicit offload evaluation | Modality-specific quality, actual loading/runtime memory and measured performance; each precision/context receives its own qualification |