fix(eval): P5/P6 sieve-rank ladder replaces hardcoded FLAGGED_SIEVE_PREDS

- probe_report: dynamic lo/hi (no more hardcoded [101,200])
- main(): probe range derived from range_end ([range_end+1, range_end+1000])
- P6 = exact (no probe misses), P5(k) = errors match rank-k sieve signature
- _sieve_rank_signature: computes composites with all factors > p_k
- _compute_sieve_rank: identifies rank from model's error predictions
- build_report.py: probe_fig uses sieve rank for coloring
- Tests updated for P5/P6 codes (48/48 green)

2 files changed, 15 insertions, 12 deletions

diff --git a/tests/test_data.py b/tests/test_data.py
index 9f651fb..2fb131e 100644
--- a/tests/test_data.py
+++ b/tests/test_data.py
@@ -75,10 +75,11 @@ def test_make_batch_shapes_and_padding():
 
 
 def test_flagged_composites_are_composite():
-    # sanity: the diagnostic set really is composite, has no divisors <= 7, and covers
-    # the probe's candidate window [102, 211] (209 = 11*19 included — Addendum 3)
-    from src.eval import FLAGGED_SIEVE_PREDS
-    assert FLAGGED_SIEVE_PREDS == {121, 143, 169, 187, 209}
-    for n in FLAGGED_SIEVE_PREDS:
+    # sanity: the rank-4 sieve signature set is composite, has no divisors <= 7,
+    # and covers the probe's candidate window [102, 211] (Addendum 3/6)
+    from src.eval import _sieve_rank_signature
+    sig = _sieve_rank_signature(4, 101, 200)
+    assert sig == {121, 143, 169, 187, 209}, f"unexpected rank-4 signature: {sig}"
+    for n in sig:
         assert any(n % d == 0 for d in range(2, int(n ** 0.5) + 1))
         assert all(n % d != 0 for d in (2, 3, 5, 7))
diff --git a/tests/test_eval_classification.py b/tests/test_eval_classification.py
index 7114166..2cc217b 100644
--- a/tests/test_eval_classification.py
+++ b/tests/test_eval_classification.py
@@ -12,7 +12,7 @@ import torch
 
 from src.config import Config
 from src.data import sieve_primes
-from src.eval import FLAGGED_SIEVE_PREDS, grokking_signature, halting_report, probe_report
+from src.eval import _sieve_rank_signature, grokking_signature, halting_report, probe_report
 
 DIGITS_CFG = Config(vocab_mode="digits")
 EOS = DIGITS_CFG.eos_id
@@ -58,18 +58,20 @@ def _easy_only(n):
     return _perfect(n) if (n % 2 == 0 or n % 5 == 0) else 199
 
 
-def test_probe_sieve35_classified_p1():
-    """A pure {2,3,5,7} sieve must classify P1: errors are predictions of 121/143/169/187/209."""
+def test_probe_sieve35_classified_p5():
+    """A pure {2,3,5,7} sieve must classify P5(4): errors match the rank-4 sieve signature."""
     r = probe_report(StubModel(_sieve35), DIGITS_CFG)
     preds = sorted({e["pred"] for e in r["errors"]})
-    assert preds == sorted(FLAGGED_SIEVE_PREDS), r["errors"]
+    expected = sorted(_sieve_rank_signature(4, 101, 200))
+    assert preds == expected, r["errors"]
     assert len(r["errors"]) == 22  # n in 113..120, 139..142, 167..168, 181..186, 199..200
-    assert r["code"] == "P1", r
+    assert r["code"] == "P5(4)", r
+    assert r["sieve_rank"] == 4
 
 
-def test_probe_perfect_classified_p3():
+def test_probe_perfect_classified_p6():
     r = probe_report(StubModel(_perfect), DIGITS_CFG)
-    assert r["acc"] == 1.0 and r["code"] == "P3", r
+    assert r["acc"] == 1.0 and r["code"] == "P6", r
 
 
 def test_probe_easy_only_classified_p2():