fix(ci,#15998): the catalog-drift check is advisory by its job name, and the docs now say so

.github/workflows/catalog-drift.yml

@@ -44,7 +44,13 @@ concurrency:
 
 jobs:
   catalog-drift:
-    name: "Notebook catalog drift (read-only)"
+    # Le marqueur `advisory` dans le NOM du job EST le contrat (#15998) : pr_gate.py
+    # classe par nom (`ADVISORY_MARKER`, regle 6) et ne lit PAS fast_lane_registry.py.
+    # Sans ce mot, la promesse "non-bloquant" de l'en-tete restait sans effet sur le
+    # gate : un echec d'infrastructure (runner, pip, generate_catalog) comptait comme
+    # un check requis et bloquait toute PR notebook/README -- observe sur #15996.
+    # Ne pas retirer le mot sans relire scripts/pr_gate.py (rule 6) et #15998.
+    name: "Notebook catalog drift (read-only, advisory)"
     # Routage #14283 tranche 3 (decision ai-01 2026-09-02) : jambe Linux
     # auto-hebergee. Le `if:` ci-dessous est la garde anti-fork exigee par
     # scripts/ci/check_self_hosted_runner_policy.py ; un `runs-on` STATIQUE la

docs/reference/catalog_markers.md

@@ -1,6 +1,6 @@
 # Catalog Markers - README Auto-Update System
 
-Source-of-truth counts driven by `COURSE_CATALOG.generated.json`. Markers in README files are expanded by `scripts/notebook_tools/expand_catalog_markers.py` and verified by CI on every PR.
+Source-of-truth counts driven by `COURSE_CATALOG.generated.json`. Markers in README files are expanded by `scripts/notebook_tools/expand_catalog_markers.py` and verified by CI on PRs that touch notebooks, series READMEs, or the catalog — the `catalog-drift.yml` workflow is filtered by `paths:`, so it does not run on every PR.
 
 ## Overview
 
@@ -75,7 +75,7 @@ python scripts/notebook_tools/expand_catalog_markers.py
 # Dry-run (show what would change)
 python scripts/notebook_tools/expand_catalog_markers.py --dry-run
 
-# Check for drift (exit 1 if stale, used by CI)
+# Check for drift (exit 1 if stale; local tool -- CI regenerates instead and never calls --check)
 python scripts/notebook_tools/expand_catalog_markers.py --check
 
 # Expand a specific file
@@ -100,12 +100,27 @@ on:
       - 'COURSE_CATALOG.generated.json'
 ```
 
-Two checks run in sequence:
-
-1. **CATALOG-STATUS marker drift** — `expand_catalog_markers.py --check` verifies all markers match the catalog
-2. **Notebook catalog drift** — `verify_catalog_readme.py` checks declared counts vs actual notebooks on disk
-
-If either check fails, the PR is blocked until markers are updated.
+Le job régénère le catalogue et les marqueurs **sur le runner** (rien n'est réécrit sur
+la branche), puis compare le résultat aux fichiers commités par un unique test
+`git diff --cached`. La dérive est remontée en **annotation `notice` uniquement**.
+
+**Ce check est advisory, non bloquant** (#15998). Le marqueur `advisory` dans le **nom**
+du job — `Notebook catalog drift (read-only, advisory)` — **est** le contrat :
+`pr_gate.py` classe les checks par nom et ne lit pas `fast_lane_registry.py`. Le job **peut rougir** : seule l'indisponibilité des métadonnées git (`rc=2` de
+`generate_catalog.py`) est absorbée en annotation `notice` ; tout autre échec (runner, checkout,
+`pip`, `generate_catalog.py` hors `rc=2`) exécute `exit "$rc"` et rend le job rouge. Mais ce
+rouge est **exclu des causes bloquantes** : le marqueur `advisory` du nom fait que `PR gate`
+le signale sans bloquer — une panne d'infrastructure est remontée, jamais bloquante pour une PR
+notebook/README (contrôle positif #16015). Le catalogue est régénéré quotidiennement
+sur `main` par `catalog-cron.yml` ; **aucune action manuelle n'est requise sur une branche
+de feature** (cf [catalog-pr-hygiene.md](../../.claude/rules/catalog-pr-hygiene.md), #2632).
+
+> Correction factuelle (2026-09-15) : cette section décrivait deux checks en séquence
+> (`expand_catalog_markers.py --check`, `verify_catalog_readme.py`) et concluait qu'un
+> échec **bloquait** la PR jusqu'à mise à jour des marqueurs. Les deux affirmations
+> étaient fausses : le workflow n'utilise pas `--check` (il régénère), ne fait appel à
+> `verify_catalog_readme.py` dans **aucun** workflow, et son job est advisory depuis
+> #15998.
 
 ## Adding Markers to a New README
 

docs/reference/ci-aggregator-rollout.md

@@ -176,7 +176,7 @@ GitHub = code pas rapport), §G.6 (audit avant merge cascade).
 | `ci / No fabricated text output in changed notebooks` | required | C.4 doc-honesty (#8052) |
 | `ci / No degenerate figure in changed notebooks` | required | GenAI rendering (cf #6541) |
 | `ci / No bare cross-dir #load in changed notebooks` | required | coupling cellule cross-répertoire |
-| `ci / Notebook catalog drift (read-only)` | advisory | catalog-pr-hygiene R1 — détecteur read-only ; le garde bloquant `catalog-pr-guard.yml` est retiré (#11012) |
+| `ci / Notebook catalog drift (read-only, advisory)` | advisory | catalog-pr-hygiene R1 — détecteur read-only ; le garde bloquant `catalog-pr-guard.yml` est retiré (#11012). Le marqueur `advisory` du nom est ce qui rend le gate conforme (#15998 : sans lui, un échec d'infra comptait comme requis) |
 | `ci / Gitleaks secret scanner` | required | secrets-hygiene |
 | `hooks-parity gate` | required | #8782 — gate qui ne peut plus rougir |
 

docs/reference/procedures-recurrentes.md

@@ -172,7 +172,7 @@ Toute PR touchant `COURSE_CATALOG.generated.json` rend les autres PRs catalog-to
 - Merger les PRs **non-catalogue d'abord**, puis les catalog-touchers **un-par-un**.
 - Conflit catalogue = l'auteur **rebase + régénère** : `python scripts/notebook_tools/generate_catalog.py --json --git-tracked-only` (parité CI = N entrées git-tracked ; `--json` nu inclut les `_output.ipynb` locaux = drift), puis `expand_catalog_markers.py`.
 - **JAMAIS** force-resolve le conflit côté coord ; **JAMAIS** force-push la branche de l'auteur.
-- Check CI **rouge "Notebook catalog drift" = NON mergeable** (propagerait le drift) → bounce à l'auteur pour re-régénérer.
+- Check CI **"Notebook catalog drift (read-only, advisory)"** : **advisory, jamais bloquant** (#15998) — le catalogue est réécrit par `catalog-cron.yml` sur `main` (#2632/#2744), une PR n'a rien à régénérer. Un drift signalé = notice informative, pas un bounce.
 - **Cascade-independence** : une PR dont le drift-check = SUCCESS **et** qui ne touche pas le catalogue est **indépendante** — ne pas la hold à tort dans la file cascade.
 
 ### Trap "APPROVED"

scripts/notebook_tools/fix_catalog_drift.py

@@ -1,8 +1,12 @@
 """One-shot fix: swap `breakdown: projects=48, Python=48` → `breakdown: Python=48, projects=48`
 in MyIA.AI.Notebooks/QuantConnect/README.md, preserving CRLF/LF unchanged.
 
-Used to unblock UNSTABLE PRs whose CI fails on `Notebook catalog drift` due to
-non-deterministic Counter.most_common() tie-break (Windows vs Linux sort order).
+Historique : ecrit pour debloquer des PRs rendues UNSTABLE par le check
+`Notebook catalog drift`, a l'epoque ou `PR gate` le comptait comme un check
+requis. Ce check est **advisory** depuis #15998 -- un drift ne bloque plus
+aucune PR, puisque le catalogue est regenere par catalog-cron.yml sur main
+(#2632/#2744). Le script reste utile comme reparation locale d'un drift de
+tie-break Counter.most_common() non deterministe (tri Windows vs Linux).
 
 Usage: python scripts/notebook_tools/fix_catalog_drift.py
 """

scripts/tests/test_pr_gate.py

@@ -1085,6 +1085,44 @@ def test_green_advisory_counts_as_a_normal_pass():
     assert ok == ["Large blob advisory (>= 10 MiB)"]
 
 
+def test_catalog_drift_job_name_carries_the_advisory_marker():
+    """#15998 -- the catalog-drift job declared itself NON-BLOCKING twice in its
+    own header (and in catalog-pr-hygiene / ci-aggregator docs), but
+    `pr_gate.py` classifies advisory by NAME (rule 6, ADVISORY_MARKER) and never
+    reads `fast_lane_registry.py`. Named "Notebook catalog drift (read-only)",
+    the job carried no marker, so any infrastructure failure (runner, pip
+    install, `generate_catalog` exit 2 on missing git metadata, cf #14831) was
+    counted as a REQUIRED check and reddened every notebook/README PR --
+    observed firsthand on #15996: "PR gate: FAIL -- failing checks: Notebook
+    catalog drift (read-only) (failure)" while every other check passed.
+
+    The repair is the marker in the job name, not a registry entry: the
+    registry is consumed by the fast lane (which would then also RUN the
+    catalog generation on every PR), whereas `is_advisory` reads the emitted
+    check-run name. Asserted over whatever the job is called today -- so a
+    future rename that keeps the marker passes, and one that drops it fails.
+    """
+    if pr_gate.yaml is None:  # pragma: no cover - PyYAML is a CI dependency
+        pytest.skip("PyYAML unavailable: cannot read the workflow")
+    wf_path = Path(pr_gate.DEFAULT_WORKFLOWS_DIR) / "catalog-drift.yml"
+    data = pr_gate.yaml.safe_load(wf_path.read_text(encoding="utf-8"))
+    job_names = pr_gate._workflow_job_names(data)
+    assert job_names, "catalog-drift.yml must declare at least one job"
+    for name in job_names:
+        assert pr_gate.is_advisory(name), (
+            "the catalog-drift job name must carry the `advisory` marker "
+            "(#15998): pr_gate.py reads the check-run name, not the header "
+            "comment, so dropping the marker silently re-arms a hard gate "
+            "against every PR touching a notebook or a series README"
+        )
+    # The defect the marker fixes must stay measurable: with the workflow name
+    # as it stands, the historical spelling is still classified BLOCKING. If
+    # this ever flips, the workflow name has absorbed the marker and the
+    # job-name invariant above stopped being the load-bearing surface.
+    wf_name = data.get("name") or ""
+    assert not pr_gate.is_advisory("Notebook catalog drift (read-only)", wf_name)
+
+
 def test_non_advisory_failure_still_blocks():
     """Guard against the fix becoming a blanket amnesty."""
     checks = [