Add production runtime standard runner

[jarcherNV]

Promote the test-only inference loop into flashdreams.runtime as run_inference_session.

The runner validates mapping compatibility before runtime creation, maps global conditioning, drives a synchronous sequential session, writes step outputs, records timing metrics, and reliably closes output/session/runtime/metrics.

Add focused production runner tests for success, validation ordering, cleanup, and declared mapping compatibility failures. Update runtime docs to mark T4 complete and point at the new runner entry point.

[greptile-apps]

Greptile Summary

The PR promotes the synchronous inference loop into the production runtime API and adds shared replay, MP4 output, and WebRTC demo infrastructure for OmniDreams.

• Validates input-mapping compatibility before runtime creation and guarantees ordered resource cleanup.
• Preserves caller-provided per-step inputs across sequential session steps.
• Adds shared demo specifications, replay and WebRTC launchers, video output support, focused tests, benchmark scenarios, and migration documentation.
• Introduces the experimental omnidreams-demo CLI and updates the workspace to support Python versions below 3.13.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains; the previously reported loss of caller-provided per-step inputs is addressed by preserving initial_inputs.step and metadata for each mapping call.

Important Files Changed

Filename	Overview
flashdreams/flashdreams/runtime/runner.py	Adds the validated sequential production runner, preserves initial per-step inputs, records timing metrics, and closes resources in deterministic order.
flashdreams/tests/test_runtime_runner.py	Covers successful execution, per-step input preservation, validation ordering, compatibility rejection, metrics, and cleanup failures.
flashdreams/flashdreams/runtime/video_output.py	Adds an MP4 output target that collects typed video step results and emits a video artifact with output metadata.
flashdreams/flashdreams/runtime/demo/replay.py	Adds shared replay orchestration that prepares an adapter scenario and delegates execution to the production runner.
flashdreams/flashdreams/runtime/demo/webrtc.py	Adds shared WebRTC runtime, manager, application, and server construction around model-specific adapters.
integrations/omnidreams/omnidreams/demo/adapter.py	Adapts OmniDreams scenario preparation and runtime creation to the shared demo API.
integrations/omnidreams/omnidreams/demo/cli.py	Introduces the experimental OmniDreams replay and WebRTC command-line entry point.
configs/omnidreams_demo_replay_benchmarks.json	Adds manual one-minute benchmark scenarios comparing the legacy and shared OmniDreams replay paths.
uv.lock	Regenerates the workspace lockfile after narrowing supported Python versions to below 3.13.

Sequence Diagram

sequenceDiagram
  participant Demo as Demo/Caller
  participant Runner as run_inference_session
  participant Mapping as InputMapping
  participant Runtime as InferenceRuntime
  participant Session as InferenceSession
  participant Output as OutputTarget
  participant Metrics as MetricsRecorder
  Demo->>Runner: adapter, config, inputs, mapping
  Runner->>Mapping: validate compatibility
  Runner->>Mapping: map global conditioning
  Runner->>Runtime: create_runtime(config)
  Runtime->>Session: start_session(mapped inputs)
  Runner->>Output: open()
  loop each StepRequest
    Session-->>Runner: next_step_request()
    Runner->>Mapping: map_step_inputs(canonical inputs, preserved step inputs)
    Runner->>Session: step(step inputs)
    Session-->>Runner: StepResult
    Runner->>Output: write(result)
    Runner->>Metrics: record timing
  end
  Runner->>Output: close()
  Runner->>Session: close()
  Runner->>Runtime: close()
  Runner->>Metrics: close()

Loading

_{Reviews (18): Last reviewed commit: "Update runtime API migration plan" | Re-trigger Greptile}