docs: add training roadmap

README.md

@@ -10,6 +10,7 @@ Docs:
 - `docs/STATUS.md`
 - `docs/ROADMAP.md`
 - `docs/REPO_FACTS.json`
+- `docs/TRAINING_ROADMAP.md`
 
 Quickstart:
 - `python -m afs_scawful datasets index`
@@ -30,3 +31,17 @@ Example `mounts.json`:
   ]
 }
 ```
+
+Training monitor (AFS Studio):
+- Use `training_monitor` in `mounts.json` or a separate `training_monitor.json` in the same config dirs.
+- Optional override: `AFS_TRAINING_MONITOR_CONFIG=/path/to/training_monitor.json`
+
+Example `training_monitor` block:
+```json
+{
+  "training_monitor": {
+    "windows_mount_path": "~/Mounts/windows-training",
+    "windows_training_dir": "D:/afs_training"
+  }
+}
+```

docs/ROADMAP.md

@@ -1,13 +1,15 @@
 # ROADMAP
 
 ## Committed
-- Minimal module layout + package stubs
-- One small utility
+- Plugin config helpers (training paths/resources + mounts)
+- Dataset registry indexing CLI
+- Resource indexing CLI
 
 ## Planned
-- Local config template
-- Example generator
+- Validator base + initial validators (ASM/C++/KG/ASAR)
+- Generator base + one small generator
+- Training monitor schema validation
 
 ## Ideas
-- Idea: Local validation hooks for training data
-- Idea: Minimal dataset manifest generator
+- Idea: Dataset manifest + QA summary reports
+- Idea: Lightweight training campaign runner (local-only)

docs/TRAINING_ROADMAP.md

@@ -0,0 +1,27 @@
+# TRAINING ROADMAP
+
+Scope: AFS Scawful training data pipelines and monitoring. Research-only.
+
+## Committed (exists now)
+- Dataset registry indexing (local)
+- Resource indexing (local)
+- Plugin config loader for training paths/resources
+
+## Planned (near-term)
+- Validator base + initial validators (ASM/C++/KG/ASAR)
+- Generator base + one small generator
+- Training monitor schema validation and config docs
+
+## Planned (mid-term)
+- Generator runner that produces dataset manifests
+- QA summary reports (counts, sizes, validation results)
+- Repeatable dataset build scripts (local-only)
+
+## Ideas (later)
+- Training campaign runner (batching + resume)
+- Per-generator metrics export (JSON)
+- Optional remote sync hooks (explicit opt-in, no auto-network)
+
+## Unknown / needs verification
+- Which generators can be safely generalized outside local workflows
+- Which validation checks are still useful vs legacy noise