record the open g2g backend gaps

Author: aaron-boxer

DESIGN_TODO.md

@@ -0,0 +1,44 @@
+# DESIGN_TODO
+
+A terse catalogue of open tasks only. Gaps on the host side of the glass2glass
+Python-element host are tracked in that repo's `DESIGN_TODO.md`, under
+"Python-element host", not here.
+
+## g2g backend coverage
+
+- **Which README pipelines run under `PYML_BACKEND=g2g` is not established
+  yet.** `PYML_BACKEND=g2g uv run pytest tests/test_pipelines.py -q
+  -p no:randomly` gives 11 passing and 67 failing: `pyml_classifier` (6
+  variants), `pyml_inference` (executorch, candle), `pyml_llm` (llamacpp),
+  `pyml_vlm`, `pyml_embedding` (clip, dinov2). Every one of the 67 failures is
+  `PIPELINE_TIMEOUT` firing at 30 s while the model loads, so the split
+  measures the card rather than g2g support. Getting the real answer needs the
+  same suite run under `gst` at the same timeout and diffed: a pipeline that
+  passes on gst and fails on g2g is a genuine gap, one that fails on both is the
+  environment. The full suite takes 10-18 minutes and wants the GPU, so do not
+  run the two backends at once on a 6 GB card.
+
+- **Eleven elements have no per-frame seam, so they cannot run on g2g at all.**
+  `alert`, `tracker`, `vad`, `clap`, `overlay_counter`, `kafkasink`,
+  `streammux`, `streamdemux`, `coalescehistory` and `llm_remote` subclass a
+  GStreamer base directly. `stablediffusion` is hosted but fills in neither
+  `process_frames` nor `process_payload`. Reparenting a family onto one of those
+  two seams in `backend/core.py` is what makes its pipelines runnable.
+  `overlay_counter` inherits `overlay`, which the launcher rewrites to g2g's
+  native `analyticsoverlay`, so the plain overlay line works regardless.
+
+- **`wavparse` will not negotiate an audio payload pipeline.**
+  `filesrc ! wavparse ! pyelement ...` fails `NoConsistentFixation` whatever the
+  caps string says. Not investigated, and it blocks every audio-in README
+  pipeline on g2g.
+
+## Elements
+
+- **`WhisperSpeechTTS.do_generate_speech` returns a `(1, n)` array**, which
+  `soundfile` rejects with `LibsndfileError: Format not recognised`, so the
+  element emits no audio. `CoquiTTS` returns 1-D and is fine. Pre-existing on
+  both backends.
+
+- **`AnomalyEngine._transform` is assigned only in `do_load_model`**, so
+  `_get_transform` raises `AttributeError` on an engine whose model never
+  loaded. Pre-existing on both backends.