test: Playwright E2E with real backend + fake TTS engine

e2e/ drives the web build (vite preview) against a per-worker uvicorn:
each Playwright worker boots its own backend on its own port with a
temp cwd, so SQLite and audio files are fully isolated and parallel-
safe. The page fixture seeds the persisted voicebox-server store with
the worker's URL before any script runs.

VOICEBOX_FAKE_TTS=1 short-circuits get_tts_backend_for_engine to a
backend that synthesizes a sine tone sized to the text — the real task
queue, SSE progress, history rows, and audio serving all run, only
inference is fake. backend/requirements-ci.txt is the slim dependency
set validated to boot the app on a CPU-only runner.

Specs: startup, settings layout, seeded profile in /voices, and
generate-to-completed-audio through the full pipeline.

backend/requirements-ci.txt

@@ -0,0 +1,25 @@
+# Minimal dependency set to boot the backend on a CPU-only CI runner.
+# No TTS/STT model libraries — inference is covered by the fake TTS
+# backend (VOICEBOX_FAKE_TTS=1). Install CPU torch first on Linux:
+#   pip install torch --index-url https://download.pytorch.org/whl/cpu
+# then: pip install -r backend/requirements-ci.txt
+
+fastapi>=0.109.0
+uvicorn[standard]>=0.27.0
+pydantic>=2.5.0
+sqlalchemy>=2.0.0
+alembic>=1.13.0
+torch>=2.2.0
+huggingface_hub>=0.20.0
+numpy
+soundfile
+python-multipart
+sse-starlette
+psutil
+requests
+httpx
+fastmcp
+librosa
+pillow
+pydub
+pedalboard