docs: update roadmap and ignore Claude settings

.claude/settings.local.json

@@ -1,20 +0,0 @@
-{
-  "permissions": {
-    "allow": [
-      "Bash(mv podcast-forge/pnpm-lock.yaml /tmp/vibepod-pnpm-lock.yaml)",
-      "Bash(git mv *)",
-      "Bash(mv /tmp/vibepod-pnpm-lock.yaml web/pnpm-lock.yaml)",
-      "Bash(git rm *)",
-      "Bash(uv lock *)",
-      "Bash(pnpm install *)",
-      "Bash(git add *)",
-      "Bash(command -v uv)",
-      "Bash(uv --version)",
-      "Bash(uv sync *)",
-      "Bash(pnpm --filter vibepod-web exec tsc --noEmit)",
-      "Bash(xargs cat *)",
-      "Bash(.venv/Scripts/python.exe -c \"import torch; print\\('torch:', torch.__version__\\); print\\('CUDA available:', torch.cuda.is_available\\(\\)\\); print\\('CUDA version:', torch.version.cuda\\)\")",
-      "Bash(nvidia-smi)"
-    ]
-  }
-}

.gitignore

@@ -24,3 +24,4 @@ web/node_modules/
 .DS_Store
 Thumbs.db
 .vscode/settings.json
+.claude/settings.local.json

roadmap.md

@@ -39,6 +39,13 @@ VibePod Studio will turn generated audio from a one-shot download into a reusabl
 - Add project save/load, autosave, and recoverable render jobs.
 - Prepare the audio pipeline for queueing longer renders outside the request lifecycle.
 
+## Later: VibeVoice Performance Research
+
+- Move the current VibePod hot-path monkey patches into the `JezzWTF/VibeVoice` fork once the feature direction has settled.
+- Add clearer generation profiling for overlapped CPU work, especially decode wait time versus total acoustic decode time.
+- Prototype batched positive/negative CFG TTS LM inference behind an opt-in flag and benchmark it against the current sequential path on CPU and CUDA.
+- Keep experimental performance work isolated from user-facing feature work unless it shows a clear speedup without audio quality regressions.
+
 ## Foundation Work Needed First
 
 - Persist generated outputs with stable IDs.