feat(ai/sidecar): online SGD+EMA trainer with replay buffer (ADR-0781)

.github/workflows/libvmaf-build-matrix.yml

@@ -1105,14 +1105,22 @@ jobs:
         if: matrix.backend == 'cuda'
         shell: cmd
         run: |
+<<<<<<< HEAD
           ninja -v -C core\build install
+=======
+          ninja -v -C libvmaf\build install
+>>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 
       - name: Build libvmaf (SYCL)
         if: matrix.backend == 'sycl'
         shell: cmd
         run: |
           call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat"
+<<<<<<< HEAD
           ninja -v -C core\build install
+=======
+          ninja -v -C libvmaf\build install
+>>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
 
       # No test step — windows-2025 has no GPU. The whole point of
       # this job is exercising the MSVC + CUDA / MSVC + oneAPI link

.github/workflows/security-scans.yml

@@ -133,11 +133,14 @@ jobs:
           languages: python
           queries: security-and-quality
           config-file: ./.github/codeql-config.yml
+<<<<<<< HEAD
       # Python analysis does not require a build step; the explicit no-op
       # below suppresses CodeQL's C++-oriented autobuild.sh which would
       # fail trying to build the old libvmaf/ path (post-rename to core/).
       - name: No-op build (Python analysis only)
         run: echo "Python-only CodeQL scan — no C++ build required"
+=======
+>>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
       - uses: github/codeql-action/analyze@9e0d7b8d25671d64c341c19c0152d693099fb5ba  # v4
         with:
           category: "/language:python"

ai/sidecar/AGENTS.md

@@ -0,0 +1,28 @@
+# ai/sidecar — Agent invariants
+
+This package implements the vmafx-node online training sidecar (ADR-0781).
+
+## Load-bearing invariants for rebase / port agents
+
+1. **Unix socket path constant** — `VMAFX_SIDECAR_SOCKET` / default
+   `/tmp/vmafx-sidecar.sock` must stay in sync between `online_trainer.py`
+   and `cmd/vmafx-node/online_feedback.go` (`feedbackSocketDefault`).  If
+   either side drifts the sidecar silently stops receiving samples.
+
+2. **Wire protocol** — the JSON envelope `{"job_id", "features", "true_score"}`
+   (Python server) and `FeedbackMessage{JobID, Features, TrueScore}` (Go client)
+   are a matched pair.  Adding a field requires updating both sides
+   atomically in the same PR.
+
+3. **ONNX opset** — `SGDEMATrainer.export_onnx()` uses `opset_version=17`
+   (ADR-0249 constraint).  Do not bump without also updating the op-allowlist
+   audit and all downstream `onnxruntime-go` consumers.
+
+4. **Replay buffer capacity default** — 10 000 samples is the ADR-0781
+   design point.  Changing it requires updating the ADR, the Helm default
+   (`sidecar.trainer.replayBufferSize`), and the docs table in
+   `docs/ai/sidecar-online-training.md`.
+
+5. **No NFL golden-data path** — this package has no connection to the
+   Netflix golden-data test fixtures or the `python/test/` assertion values.
+   Changes here cannot affect those tests.

ai/sidecar/__init__.py

@@ -0,0 +1,19 @@
+# SPDX-License-Identifier: BSD-3-Clause-Plus-Patent
+# Copyright 2026 Lusoris
+#
+# ai/sidecar — vmafx-node online training sidecar package.
+#
+# ADR-0781: sidecar online training — SGD + EMA + replay buffer.
+
+from .online_trainer import OnlineTrainer, run_server
+from .replay_buffer import ReplayBuffer, Sample
+from .sgd_ema import SGDEMAConfig, SGDEMATrainer
+
+__all__ = [
+    "OnlineTrainer",
+    "ReplayBuffer",
+    "SGDEMAConfig",
+    "SGDEMATrainer",
+    "Sample",
+    "run_server",
+]