Fix scission palier inference and create output dirs

**Motivations:**
- Make certificates reproducible when CSV columns do not encode the palier
- Avoid FileNotFoundError when writing certificates into new folders
- Reuse scission in the local H6 generator to avoid duplicated certificate logic

**Root causes:**
- palier inference relied on max residue value when the class column was generic
- scission assumed output directories already exist
- empty CSV fields were coerced to 0

**Correctifs:**
- Infer palier from explicit columns (palier/m) or filename, keep heuristic fallback
- Create parent directory for output JSON
- Skip empty class/sister values instead of adding residue 0

**Evolutions:**
- Use collatz_scission for certificate generation in local H6 artefacts generator

**Pages affectées:**
- applications/collatz/collatz_k_scripts/collatz_scission.py
- applications/collatz/collatz_k_scripts/collatz_generate_local_h6_artefacts.py
- docs/fixKnowledge/collatz_scission_palier_inference_and_output_dirs.md

applications/collatz/collatz_k_scripts/collatz_generate_local_h6_artefacts.py

@@ -38,6 +38,7 @@ from pathlib import Path
 from collatz_k_core import A_k
 from collatz_k_fusion import build_fusion_clauses
 from collatz_k_pipeline import load_state_map_60
+from collatz_scission import run_scission
 
 
 @dataclass(frozen=True)
@@ -58,14 +59,22 @@ class LocalH6StateResult:
     artefacts_dir: str
 
 
-def _write_certificate_json(path: Path, clauses: set[int], covered: set[int], palier: int) -> None:
+def _read_certificate_json(cert_path: Path) -> tuple[set[int], set[int], int]:
-    path.parent.mkdir(parents=True, exist_ok=True)
+    obj = json.loads(cert_path.read_text(encoding="utf-8"))
-    obj = {
+    if not isinstance(obj, dict):
-        "clauses": sorted(clauses),
+        raise ValueError(f"Invalid certificate JSON (expected object): {cert_path}")
-        "covered": sorted(covered),
+    clauses_raw = obj.get("clauses", [])
-        "palier": palier,
+    covered_raw = obj.get("covered", [])
-    }
+    palier_raw = obj.get("palier", 0)
-    path.write_text(json.dumps(obj, indent=2), encoding="utf-8")
+
+    if not isinstance(clauses_raw, list) or not all(isinstance(x, int) for x in clauses_raw):
+        raise ValueError(f"Invalid certificate JSON clauses list: {cert_path}")
+    if not isinstance(covered_raw, list) or not all(isinstance(x, int) for x in covered_raw):
+        raise ValueError(f"Invalid certificate JSON covered list: {cert_path}")
+    if not isinstance(palier_raw, int):
+        raise ValueError(f"Invalid certificate JSON palier: {cert_path}")
+
+    return set(clauses_raw), set(covered_raw), palier_raw
 
 
 def _invert_residue_to_state(res_to_state: dict[int, int]) -> dict[int, list[int]]:
@@ -110,13 +119,8 @@ def _write_candidates_d8_csv(path: Path, residues: list[int], palier: int) -> li
 
 
 def _covered_from_cert(cert_path: Path) -> set[int]:
-    obj = json.loads(cert_path.read_text(encoding="utf-8"))
+    _, covered, _ = _read_certificate_json(cert_path)
-    if not isinstance(obj, dict) or "covered" not in obj:
+    return covered
-        raise ValueError(f"Invalid certificate JSON: {cert_path}")
-    covered = obj["covered"]
-    if not isinstance(covered, list) or not all(isinstance(x, int) for x in covered):
-        raise ValueError(f"Invalid certificate JSON covered list: {cert_path}")
-    return set(covered)
 
 
 def _compute_fusion_rows_until_covered(
@@ -293,9 +297,12 @@ def generate_for_state(
         d8_covered_set: set[int] = set()
         cert_paths: list[Path] = []
         if cand_d8:
-            shift = 1 << (palier - 1)
+            run_scission(str(candidats_d8_path), str(certificat_d8_path))
-            d8_covered_set = set(cand_d8) | {n ^ shift for n in cand_d8}
+            _, d8_covered_set, cert_palier = _read_certificate_json(certificat_d8_path)
-            _write_certificate_json(certificat_d8_path, set(cand_d8), d8_covered_set, palier)
+            if cert_palier != palier:
+                raise ValueError(
+                    f"Unexpected D8 certificate palier for E{state_id}: got 2^{cert_palier}, expected 2^{palier}"
+                )
             cert_paths.append(certificat_d8_path)
 
         uncovered_after_d8 = sorted(set(L) - d8_covered_set)
@@ -322,16 +329,15 @@ def generate_for_state(
         fusion_covered_set: set[int] = set()
         fusion_clauses = 0
         fusion_covered = 0
-        fusion_hit_classes = {
+        if fusion_rows_list and fusion_cert_path is not None and fusion_csv_path is not None:
-            int(row["classe_mod_2^m"])
+            run_scission(str(fusion_csv_path), str(fusion_cert_path))
-            for row in fusion_rows_list
+            fusion_clauses_set, fusion_covered_set, cert_palier = _read_certificate_json(fusion_cert_path)
-            if "classe_mod_2^m" in row and str(row["classe_mod_2^m"]).strip()
+            if cert_palier != palier:
-        }
+                raise ValueError(
-        if fusion_hit_classes and fusion_cert_path is not None:
+                    f"Unexpected fusion certificate palier for E{state_id}: got 2^{cert_palier}, expected 2^{palier}"
-            fusion_covered_set = set(fusion_hit_classes)
+                )
-            _write_certificate_json(fusion_cert_path, set(fusion_hit_classes), fusion_covered_set, palier)
+            fusion_clauses = len(fusion_clauses_set)
-            fusion_clauses = len(fusion_hit_classes)
+            fusion_covered = len(fusion_covered_set)
-            fusion_covered = len(fusion_hit_classes)
             cert_paths.append(fusion_cert_path)
 
         union_covered = d8_covered_set | fusion_covered_set

applications/collatz/collatz_k_scripts/collatz_scission.py

@@ -26,12 +26,54 @@ def _find_column(row: dict, *candidates: str) -> str | None:
     return None
 
 
-def infer_palier(rows: list[dict], classe_col: str | None) -> int:
+def _try_parse_int(value: object) -> int | None:
-    """Infer modulus power m from column name or max value."""
+    if value is None:
+        return None
+    if isinstance(value, int):
+        return value
+    if isinstance(value, str):
+        s = value.strip()
+        if not s:
+            return None
+        try:
+            return int(s)
+        except ValueError:
+            return None
+    return None
+
+
+def infer_palier(rows: list[dict], classe_col: str | None, csv_path: Path | None = None) -> int:
+    """
+    Infer modulus power m.
+
+    Priority order:
+    - explicit numeric column 'palier' (or 'm' used as exponent in some CSVs)
+    - class column name containing '2^<m>' (e.g. 'classe_mod_2^27')
+    - filename containing 'palier2p<m>'
+    - fallback heuristic from max class value (legacy; not reliable when values are sparse)
+    """
+    if rows:
+        pal_col = _find_column(rows[0], "palier")
+        if pal_col:
+            v = _try_parse_int(rows[0].get(pal_col))
+            if v is not None and v > 0:
+                return v
+        m_col = _find_column(rows[0], "m", "modulus_power")
+        if m_col:
+            v = _try_parse_int(rows[0].get(m_col))
+            if v is not None and v > 0:
+                return v
+
     if classe_col and ("2^" in classe_col or "2^" in str(classe_col)):
         m = re.search(r"2\^(\d+)", classe_col)
         if m:
             return int(m.group(1))
+
+    if csv_path is not None:
+        m2 = re.search(r"palier2p(\d+)", str(csv_path))
+        if m2:
+            return int(m2.group(1))
+
     if rows and classe_col:
         try:
             vals = [int(r.get(classe_col, 0) or 0) for r in rows if r.get(classe_col)]
@@ -48,6 +90,9 @@ def infer_palier(rows: list[dict], classe_col: str | None) -> int:
 
 def run_scission(csv_path: str, out_json_path: str) -> None:
     """Read CSV, extract clauses and covered set, write JSON certificate."""
+    out_path = Path(out_json_path)
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+
     rows: list[dict] = []
     with Path(csv_path).open("r", encoding="utf-8") as f:
         reader = csv.DictReader(f)
@@ -56,7 +101,7 @@ def run_scission(csv_path: str, out_json_path: str) -> None:
 
     if not rows:
         cert = {"clauses": [], "covered": [], "palier": 0}
-        Path(out_json_path).write_text(json.dumps(cert, indent=2), encoding="utf-8")
+        out_path.write_text(json.dumps(cert, indent=2), encoding="utf-8")
         print(f"Wrote {out_json_path} (empty)")
         return
 
@@ -69,28 +114,32 @@ def run_scission(csv_path: str, out_json_path: str) -> None:
     for r in rows:
         if classe_col:
             try:
-                c = int(r.get(classe_col, 0) or 0)
+                raw = r.get(classe_col)
-                clauses.append(c)
+                if raw is not None and str(raw).strip():
-                covered.add(c)
+                    c = int(raw)
+                    clauses.append(c)
+                    covered.add(c)
             except (ValueError, TypeError):
                 pass
         if soeur_col:
             try:
-                s = int(r.get(soeur_col, 0) or 0)
+                raw = r.get(soeur_col)
-                covered.add(s)
+                if raw is not None and str(raw).strip():
+                    s = int(raw)
+                    covered.add(s)
             except (ValueError, TypeError):
                 pass
 
     clauses = sorted(set(clauses))
     covered_list = sorted(covered)
-    palier = infer_palier(rows, classe_col)
+    palier = infer_palier(rows, classe_col, csv_path=Path(csv_path))
 
     cert = {
         "clauses": clauses,
         "covered": covered_list,
         "palier": palier,
     }
-    Path(out_json_path).write_text(json.dumps(cert, indent=2), encoding="utf-8")
+    out_path.write_text(json.dumps(cert, indent=2), encoding="utf-8")
     print(f"Wrote {out_json_path}: {len(clauses)} clauses, {len(covered_list)} covered, palier 2^{palier}")
 
 

docs/fixKnowledge/collatz_scission_palier_inference_and_output_dirs.md

@@ -0,0 +1,45 @@
+# collatz_scission palier inference and output directory creation
+
+## Problem
+
+The helper script `applications/collatz/collatz_k_scripts/collatz_scission.py` can produce incorrect certificates in two cases:
+
+- **Incorrect `palier` inference**: when the CSV class column is named generically (e.g. `classe_mod_2^m`) and the covered classes are sparse / small (e.g. values \(<2^8\) while the target modulus is \(2^{13}\)), `palier` was inferred from the maximum class value. This yields a wrong modulus power.
+- **Missing output directories**: `run_scission()` writes `out_json_path` without creating parent directories, which can raise `FileNotFoundError` when callers pass a new path under a non-existing folder.
+
+## Root cause
+
+- `infer_palier()` only supported:
+  - parsing `2^m` from the class column name, or
+  - a fallback heuristic based on the maximum covered residue value.
+  This heuristic is not reliable when the class column name does not encode the modulus power.
+- `run_scission()` assumed the output directory exists.
+
+## Corrective actions
+
+- **Prefer explicit palier columns**:
+  - If the CSV contains a numeric `palier` column, use it.
+  - If the CSV contains a numeric `m` / `modulus_power` column (used as exponent in some pipelines), use it.
+- **Fallback from filename**: parse `palier2p<m>` from the CSV path when available.
+- **Keep legacy fallback**: keep the max-value heuristic as a last resort.
+- **Create output directories**: ensure `out_json_path.parent` exists before writing.
+- **Do not add spurious residue 0**: skip empty strings instead of coercing to 0 when parsing the class / sister columns.
+
+## Impact
+
+- Certificates generated via `collatz_scission.py` now carry a `palier` that matches the CSV’s intended modulus power when the CSV provides it (or when the filename encodes it).
+- Callers can write certificates to new directories without pre-creating them.
+
+## Analysis modalities
+
+- For any certificate JSON, verify:
+  - `palier` matches the intended modulus power \(2^m\),
+  - `clauses` and `covered` sets do not contain a spurious `0`,
+  - directory creation does not fail when writing under a fresh path.
+
+## Deployment
+
+- No environment changes are required.
+- The fix is local to:
+  - `applications/collatz/collatz_k_scripts/collatz_scission.py`
+