P3: normalization fit (1.2.0 and 1.1.2), apply, polar, simurg normalize

- signal/normalization/numerics.py: vendor percentile, trimmean, MAD sigma
  (old/new), bisquare robust fits, float32/float64 typing and sequential
  sums reproduced exactly (ported from gtc2vcf, MIT)
- signal/normalization/fit.py: outlier removal, offset, rotation, shear,
  scale, nn12/rr12, single-channel model for small Infinium I bins,
  exact and binned nearest-neighbour sweeps; vendor quirks marked QUIRK
- signal/normalization/apply.py: affine transform and theta/R
- pipeline/normalize.py + CLI `simurg normalize`
- oracle GTCs rebuilt with -ffp-contract=off (aarch64 gcc fuses a*b+c by
  default, shifting the last bits); the old FMA oracle differed by <=4 ULP

Result: transforms bit-identical to the oracle for all 36 GSA samples x
48 bins, for both normalization versions (3456/3456 bins, 0 ULP).

tests/oracle/test_p2_raw_xy.py

@@ -6,12 +6,12 @@ from __future__ import annotations
 
 import numpy as np
 import pytest
+from tests.conftest import GSA_BPM, GSA_IDATS, ORACLE_GTCS, need
 
 from simurg.formats.bpm import read_bpm
 from simurg.formats.gtc import read_gtc
 from simurg.formats.idat import read_idat
 from simurg.signal.raw_xy import LocusMapCache, assemble_raw_xy, check_coverage
-from tests.conftest import GSA_BPM, GSA_IDATS, ORACLE_GTCS, need
 
 pytestmark = pytest.mark.data
 

tests/oracle/test_p3_normalization.py

@@ -0,0 +1,45 @@
+"""Phase P3 acceptance on P0 data: normalization transforms fitted from raw intensities are bit-identical to the
+transforms stored in oracle GTCs (gtc2vcf built without FMA), for normalization 1.2.0 and 1.1.2.
+The same check against vendor GTCs runs in scripts/validate_p1_large.py --p3.
+"""
+
+from __future__ import annotations
+
+import numpy as np
+import pytest
+from tests.conftest import GSA_BPM, ORACLE_GTCS, ORACLE_GTCS_GT2, need
+
+from simurg.formats.bpm import read_bpm
+from simurg.formats.gtc import read_gtc
+from simurg.signal.normalization.fit import fit_transforms
+
+pytestmark = pytest.mark.data
+
+FIELDS = ("offset_x", "offset_y", "scale_x", "scale_y", "shear", "theta")
+
+
+def as_bits(xf) -> np.ndarray:  # type: ignore[no-untyped-def,type-arg]
+    v = [getattr(xf, f) for f in FIELDS] + [xf.nn12, xf.rr12, xf.taa, xf.tbb]
+    return np.array(v, dtype=np.float32).view(np.int32)
+
+
+def ref_bits(rec) -> np.ndarray:  # type: ignore[no-untyped-def,type-arg]
+    v = [rec[f] for f in FIELDS] + list(rec["reserved"][2:6])
+    return np.array(v, dtype=np.float32).view(np.int32)
+
+
+@pytest.mark.parametrize(
+    ("folder", "version"), [(ORACLE_GTCS, "1.2.0"), (ORACLE_GTCS_GT2, "1.1.2")], ids=["1.2.0", "1.1.2"]
+)
+def test_transforms_bit_identical(folder, version) -> None:  # type: ignore[no-untyped-def]
+    need(GSA_BPM, folder)
+    m = read_bpm(GSA_BPM)
+    bins = m.norm_bin_ids()
+    gtcs = sorted(folder.glob("*.gtc"))
+    assert len(gtcs) >= 18
+    for g in gtcs:
+        ref = read_gtc(g)
+        mine = fit_transforms(ref.raw_x, ref.raw_y, bins, version)
+        assert len(mine) == len(ref.normalization_transforms)
+        for k, (a, b) in enumerate(zip(mine, ref.normalization_transforms, strict=True)):
+            assert np.array_equal(as_bits(a), ref_bits(b)), f"{g.name} bin {k}"