Refactor: reorganize modules + add profile-driven per-basis SCF for pooled CorrNet training

Structural refactors:
- Move deepks/io/input/ → deepks/config/ (packager, validator, defaults, loader, etc.)
- Flatten deepks/orchestration/scheduler/job/ → deepks/orchestration/scheduler/
- Move ABACUS iterate/scf workflow files to deepks/workflows/iterate/abacus/ and deepks/workflows/scf/abacus/
- Move runtime reporting to deepks/io/reporting.py

New feature: profile-driven per-basis SCF for non-hierarchical recipes
- task_params.py: add resolve_scf_profile_levels() — drives per-basis SCF Sequence
  from physics.backend.profiles for any recipe (not gated on hierarchical-regression)
- prepare.py: scf_levels = hierarchical_levels or profile_levels; use for main+init SCF
- packager.py: use_profile_scf path sets pooled train data paths level.00/01/02
- validator.py: relax global orb_files requirement when profiles present; validate per-profile

Enables training a single shared CorrNet on pooled sz/dzp/tzdp data in one iterate,
testing basis-independence without the additive hierarchical stack.

Author: ErjieWu

.codex

@@ -1,6 +1,6 @@
 """Unified input configuration module for DeePKS."""
 
-from .config import normalize_config
+from .normalize import normalize_config
 from .docs import render_input_parameter_doc
 from .defaults import get_default_config
 from .loader import load_config

deepks/io/input/__init__.py deepks/config/__init__.pydeepks/io/input/__init__.py renamed to deepks/config/__init__.py

@@ -327,7 +327,7 @@ def render_input_parameter_doc():
     lines = [
         "# DeePKS input parameter reference",
         "",
-        "> Generated from `deepks/io/input/docs.py`. Do not edit this file manually.",
+        "> Generated from `deepks/config/docs.py`. Do not edit this file manually.",
         "",
         "## Preferred schema",
         "",

deepks/io/input/defaults.py deepks/config/defaults.pydeepks/io/input/defaults.py renamed to deepks/config/defaults.py

@@ -333,6 +333,8 @@ def pack_child(child_type, child_config):
         )
         hierarchical_train_groups = data.get("train", []) if isinstance(data.get("train"), list) else []
         hierarchical_test_groups = data.get("test", []) if isinstance(data.get("test"), list) else []
+        scf_profiles = backend.get("profiles") if isinstance(backend.get("profiles"), list) else []
+        use_profile_scf = bool(scf_profiles) and not hierarchical_levels
 
         main_scf = {
             "recipe": deepcopy(main_config.get("recipe")),
@@ -393,8 +395,17 @@ def pack_child(child_type, child_config):
             resolved_terms = _resolve_hierarchical_terms(ml)
             if resolved_terms:
                 main_train.setdefault("ml", {}).setdefault("objective", {})["terms"] = resolved_terms
+        elif use_profile_scf:
+            main_train["data"]["train"] = [
+                f"../00.scf/level.{i:02d}/data_train/*" for i in range(len(scf_profiles))
+            ]
         if data.get("test") is not None:
-            main_train["data"]["test"] = "data_test/*"
+            if use_profile_scf:
+                main_train["data"]["test"] = [
+                    f"../00.scf/level.{i:02d}/data_test/*" for i in range(len(scf_profiles))
+                ]
+            else:
+                main_train["data"]["test"] = "data_test/*"
         if isinstance(runtime.get("io"), dict):
             main_train["runtime"]["io"] = deepcopy(runtime["io"])
         if child_proj_basis:

deepks/io/input/dispatcher.py deepks/config/dispatcher.pydeepks/io/input/dispatcher.py renamed to deepks/config/dispatcher.py

@@ -171,11 +171,30 @@ def _validate_iterate_config(config):
         if n_iter < 0:
             raise ValueError(f"'n_iter' must be non-negative, got {n_iter}")
 
+    physics = config.get("physics") if isinstance(config.get("physics"), dict) else {}
+    backend = physics.get("backend") if isinstance(physics.get("backend"), dict) else {}
+    scf_profiles = backend.get("profiles") if isinstance(backend.get("profiles"), list) else []
     backend_name = _validate_scf_backend_name(config)
     if backend_name == 'pyscf':
         _validate_pyscf_block(config)
     else:
-        _validate_abacus_block(config, require_files=(recipe_name != HIERARCHICAL_REGRESSION_RECIPE_NAME))
+        # Per-basis SCF profiles supply orb_files individually, so the global
+        # physics.backend.input.orb_files requirement is relaxed when present.
+        require_files = recipe_name != HIERARCHICAL_REGRESSION_RECIPE_NAME and not scf_profiles
+        _validate_abacus_block(config, require_files=require_files)
+        # Profile-driven per-basis SCF (non-hierarchical): each profile must
+        # carry its own orb_files (or an input_template). The hierarchical
+        # recipe validates its own profiles in _validate_hierarchical_iterate_config.
+        if recipe_name != HIERARCHICAL_REGRESSION_RECIPE_NAME:
+            for i, prof in enumerate(scf_profiles):
+                prof_input = prof.get("input", {}) if isinstance(prof, dict) else {}
+                has_orb = isinstance(prof_input, dict) and prof_input.get("orb_files")
+                has_template = isinstance(prof, dict) and prof.get("input_template") is not None
+                if not (has_orb or has_template):
+                    raise ValueError(
+                        f"physics.backend.profiles[{i}] requires 'input.orb_files' "
+                        "(or 'input_template') for per-basis SCF"
+                    )
 
 
 def _validate_hierarchical_iterate_config(config):