python-math #45: test(delphi): FLI-cold_start repful comparison → xfail (prod-blob Q12) + s7 Phase 4 journal

delphi/docs/CLJ-PARITY-FIXES-JOURNAL.md

@@ -4262,3 +4262,29 @@ Phase 4 re-record therefore = private goldens (--include-local) +
 optionally recording public ones; VERIFY against the battery's certified
 clj recordings BEFORE recording (never blind). Recorder:
 scripts/regression_recorder.py.
+
+### Phase 4 — goldens: verify-then-re-record (s7 cont.)
+
+Drift check (`regression_comparer.py --include-local`): 7/7 datasets fail
+vs the pre-collapse goldens. Read the FLI detail per the never-blind rule
+— the drift is EXACTLY the two expected legacy families and nothing else:
+(1) blob shape (folded columnar base-clusters, lastModTimestamp/mod-out/
+mod-in emission, legacy consensus/repness shapes — _apply_legacy_blob_shape
+now unconditional); (2) PCA warm-start numerics in the rank-deficient
+component tail (comps[1]/proj second-axis jitter at tiny magnitudes).
+
+Verification argument for re-recording (the "certified before recording"
+evidence): the engine writing the new goldens is the SAME TREE the battery
+just certified bit-exact against the Clojure oracle on these SAME 7
+datasets (20/20 MATCH, two runs — post-collapse and post-refactor). The
+observed drift families match the collapse's documented semantics 1:1;
+no third family observed. Goldens are LOCAL-ONLY artifacts (zero
+git-tracked golden_snapshot.json), so the re-record is evidenced by the
+comparer passing + this entry, not by committed files. Public datasets
+(vw/biodiversity) had NO goldens at all — being recorded for the first
+time in this worktree.
+
+PR #2673 review agent verdict: CLEAN — split proven behavior-preserving
+(including an equivalence proof of the counts_df.empty seam), blob test
+non-vacuous (45+42 entries), tolerances justified against Clojure's
+`(float repness-test)` cast at repness.clj:187.

delphi/tests/test_discrepancy_fixes.py

@@ -1077,12 +1077,19 @@ def test_repful_classification_boundary(self):
     def test_repful_matches_clojure_blob(self, request, conv, clojure_blob, dataset_name):
         """repful-for (Clojure) vs repful (Python) for shared rep comments.
 
-        Gates on 9/11 variants since the gid label-swap fix (2026-07-05
+        Gates on most variants since the gid label-swap fix (2026-07-05
         removal of the group size re-sort). Residual known-bad: two
         incremental variants with deeper trajectory divergence
         (pakistan-incremental: Clojure blob PCA computed on a comment
         subset; vw-incremental: in-conv trajectory divergence) — deferred
-        to the sequential-parity work.
+        to the sequential-parity work — plus FLI-cold_start since the mode
+        collapse (2026-07-27): the PROD blob carries Clojure's UNSEEDED
+        cold-start PCA (Q12, #2661) and the collapse made the
+        Clojure-faithful legacy kmeans the only cold-tick path, so the
+        selection sets no longer intersect that particular random draw.
+        The battery certifies FLI end-to-end against pinned-cold-start
+        Clojure recordings (20/20 MATCH), which supersedes this prod-blob
+        comparison for that variant.
         """
         if request.node.callspec.id in ('vw-incremental', 'pakistan-incremental'):
             request.applymarker(pytest.mark.xfail(
@@ -1091,6 +1098,13 @@ def test_repful_matches_clojure_blob(self, request, conv, clojure_blob, dataset_
                 reason="residual incremental trajectory divergence (gid "
                        "label swap fixed 2026-07-05; sequential-parity "
                        "work)"))
+        if request.node.callspec.id == 'FLI-cold_start':
+            request.applymarker(pytest.mark.xfail(
+                strict=False,
+                reason="prod blob Q12 unseeded cold-start PCA (#2661) vs "
+                       "the collapse's legacy-kmeans cold tick — zero "
+                       "shared selections; battery certifies FLI against "
+                       "pinned-cold-start recordings instead"))
         clojure_repness = clojure_blob.get('repness', {})
         if not clojure_repness:
             pytest.skip("No repness in Clojure blob")