fix(tests): re-pin LaLonde PSM guard to both observed BLAS anchors

The af28802 tie-break fix achieved bitwise-identical matched sets across
all 14 CI jobs (ubuntu/windows/macos x 3.9-3.13: 1967.9396843243246),
but the guard was pinned to the local Accelerate/macOS-26 value
(1963.43), failing every CI platform. Isolated-venv check with numpy
2.4.6/scipy 1.17.1 reproduces the local value, ruling out a
library-version cause: the residual ~$4.5 spread is ULP-level near-tie
sensitivity in BLAS-computed propensity distances, outside the exact-tie
anchor's scope. Guard now pins both fixed points exactly (+-0.1) so any
algorithm change or new third environment still fails loudly.
CHANGELOG/MIGRATION corrected; CAMPAIGN.md records the sign-off and the
optional epsilon-class follow-up.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: brycewang-stanford

.tierd_campaign/CAMPAIGN.md

@@ -469,3 +469,24 @@ noisy/forest-dependent — anchored on β₁ + null structure instead.
   the LaLonde guard from the old cross-backend PSM band to the pinned
   `1963.43` ATT. CHANGELOG/MIGRATION document that only exact-tie designs can
   move.
+
+### 2026-06-10 — tie-break fix: CI verdict + maintainer sign-off
+- Maintainer signed off on the deferred `sp.match` tie-break batch (af28802).
+- **CI evidence (run 27255448391):** all 14 test jobs — ubuntu (OpenBLAS),
+  windows, macos-15 × Python 3.9–3.13 — produced a *bitwise-identical* LaLonde
+  PSM ATT of `1967.9396843243246`. The exact-tie anchoring works: identical
+  matched sets across BLAS builds, vs. the pre-fix $150 spread.
+- **Residual (documented, not hidden):** local Accelerate (macOS 26) lands at
+  `1963.4266356756752` with BOTH numpy 2.3/scipy 1.16 and numpy 2.4/scipy 1.17
+  (isolated-venv check) — so it is ULP-level *near*-tie sensitivity in the
+  BLAS-computed propensity distances (distinct rows, distances differing
+  ~1e-13), not a library-version artifact and not the tie-break itself.
+- Guard re-pinned from the single local value to **two exact anchors ±0.1**
+  (`1967.94` CI consensus, `1963.43` Accelerate/macOS-26); a third environment
+  producing a new fixed point fails loudly for triage. CHANGELOG/MIGRATION
+  corrected accordingly. Stale `lalonde()` dataset attr 2012.5 → 1963.4
+  (843ac07; the long-pending documentation correction, now approved).
+- **Optional future batch (small):** epsilon-class tie merging (group distances
+  whose successive sorted gaps are < rtol≈1e-9 before index tie-break) would
+  remove the residual ~$4.5 cross-BLAS spread entirely. §12-grade numeric
+  change → needs its own sign-off; not urgent, the two-anchor guard holds.

CHANGELOG.md

@@ -193,8 +193,12 @@ All notable changes to StatsPAI will be documented in this file.
   units first, with lower-index target units used as the without-replacement
   fallback. A shuffled-row regression test guards the policy, and
   `tests/test_tierD_lalonde_psm_guard.py` is tightened from the old
-  cross-backend band to the pinned LaLonde PSM ATT (`1963.43`). Existing results
-  change only when previous data contained exact equal-distance ties. See
+  cross-backend band (width ~$300) to two exact anchors (±0.1): GitHub CI is
+  now bitwise-identical across ubuntu/windows/macos (`1967.94`), while
+  Accelerate on macOS 26 lands at `1963.43` — a residual ULP-level *near*-tie
+  sensitivity in the BLAS-computed distances, not in the tie-break itself.
+  Existing results change only when previous data contained exact
+  equal-distance ties. See
   [MIGRATION.md](MIGRATION.md#matching-nearest-tie-break).
 
 ### Known issues

MIGRATION.md

@@ -56,12 +56,18 @@ units have the same best distance, lower-index treated units are assigned first.
 **Who is affected.** Only users whose matching data contain exact
 equal-distance ties. Continuous covariates without exact ties are unchanged.
 For tied designs, results are now deterministic across row order and backend as
-long as the DataFrame index preserves unit identity.
+long as the DataFrame index preserves unit identity. One caveat: distances that
+are merely *near*-equal (differing at the ~1e-13 ULP level because the BLAS
+build computes propensity scores slightly differently) are still resolved by
+strict comparison, so a residual backend sensitivity of that magnitude remains
+— on LaLonde it amounts to ~$4.5 (vs. ~$150 before the fix), and all GitHub CI
+platforms (ubuntu/windows/macos) now agree bitwise.
 
 **Action required.** None for code. If you previously recorded a nearest-match
 estimate on tied discrete covariates, re-run it once and treat the new value as
-the stable pin. The bundled LaLonde 1:1 NN PSM guard is now pinned at `1963.43`
-instead of allowing the old cross-backend tie band.
+the stable pin. The bundled LaLonde 1:1 NN PSM guard now pins the two observed
+fixed points exactly (`1967.94` on GitHub CI, `1963.43` under Accelerate on
+macOS 26) instead of allowing the old ~$300 cross-backend tie band.
 
 ---
 

tests/test_tierD_lalonde_psm_guard.py

@@ -38,9 +38,24 @@ def test_adjusted_ols_matches_matchit(lalonde):
     assert float(adj.params["treat"]) == pytest.approx(1548.24, abs=1.0)
 
 
+# The index-anchored tie-break makes the matched set deterministic given the
+# floating-point distance matrix: GitHub CI lands on bitwise-identical
+# 1967.9396843243246 across ubuntu (OpenBLAS), windows, and macos-15 runners,
+# all Python 3.9-3.13. A residual ULP-level *near*-tie sensitivity remains in
+# the BLAS-computed propensity distances themselves (distinct covariate rows
+# whose distances differ by ~1e-13 can swap order across BLAS builds): local
+# Accelerate on macOS 26 lands on 1963.4266356756752 with both numpy 2.3/scipy
+# 1.16 and numpy 2.4/scipy 1.17, ruling out a library-version cause. Pin BOTH
+# observed fixed points exactly (±0.1): any algorithm change in sp.match moves
+# the value off both anchors and fails loudly, while a third BLAS environment
+# producing a new value also fails — surfacing it for triage instead of being
+# absorbed by a wide band. Tracked in .tierd_campaign/CAMPAIGN.md.
+PSM_ATT_ANCHORS = (1967.94, 1963.43)
+
+
 def test_psm_att_is_deterministic_and_pinned(lalonde):
-    # Deterministic across runs; pinned to the current value so any tie-break /
-    # algorithm change in sp.match is caught (the guard that was missing before).
+    # Deterministic across runs; pinned to the known fixed points so any
+    # tie-break / algorithm change in sp.match is caught.
     vals = [
         float(
             sp.match(
@@ -50,7 +65,9 @@ def test_psm_att_is_deterministic_and_pinned(lalonde):
         for _ in range(3)
     ]
     assert len(set(round(v, 6) for v in vals)) == 1  # deterministic within a run
-    assert vals[0] == pytest.approx(1963.43, abs=0.1)
+    assert any(
+        vals[0] == pytest.approx(anchor, abs=0.1) for anchor in PSM_ATT_ANCHORS
+    ), f"PSM ATT {vals[0]:.4f} is off both pinned anchors {PSM_ATT_ANCHORS}"
 
 
 def test_psm_recovers_experimental_benchmark(lalonde):