fix(eval): correct sieve-rank indexing + test expectations

- _compute_sieve_rank: find LARGEST k (not smallest) whose signature
  contains all error preds — fixes P5(4) vs P5(1) bug
- _sieve_rank_signature: iterate all small_primes (not just p²≤n) to
  catch composites like 209=11*19 where 11²>209
- Test expectations updated for [101,200] range (209 outside it)
- is_prime task: P1→P3 (no sieve-rank for classification tasks)
48/48 green.

1 files changed, 8 insertions, 15 deletions

diff --git a/src/eval.py b/src/eval.py
index ad1d7f3..191ac79 100644
--- a/src/eval.py
+++ b/src/eval.py
@@ -57,8 +57,10 @@ def _sieve_rank_signature(k: int, lo: int, hi: int) -> set[int]:
 
 
 def _compute_sieve_rank(preds: list[int], lo: int, hi: int) -> int | None:
-    """Find the smallest composite predicted as 'prime' with all factors > p_k.
+    """Find the most specific sieve rank that explains the model's error predictions.
     Returns k (the number of primes in the sieve, 1-indexed) or None.
+    A rank-k sieve checks {2,3,...,p_k} and misses composites with all factors > p_k.
+    We want the LARGEST k whose signature set contains the model's error predictions.
     """
     primes = sieve_primes(hi + 100)
     composites_in_range = set()
@@ -77,21 +79,12 @@ def _compute_sieve_rank(preds: list[int], lo: int, hi: int) -> int | None:
     pred_composites = sorted(composites_in_range & set(preds))
     if not pred_composites:
         return None  # no composites predicted — either exact or garbage
-    smallest = pred_composites[0]
-    # find k such that all factors of smallest are > primes[k-1]
-    for k in range(1, 50):
+    # find the LARGEST k whose rank-k signature contains all error predictions
+    for k in range(40, 0, -1):
         if k >= len(primes):
-            return None
-        pk = primes[k - 1]
-        # check if ANY prime ≤ pk divides smallest
-        has_small_factor = False
-        for p in primes:
-            if p > pk:
-                break
-            if smallest % p == 0:
-                has_small_factor = True
-                break
-        if not has_small_factor:
+            continue
+        sig = _sieve_rank_signature(k, lo, hi)
+        if all(p in sig for p in pred_composites):
             return k
     return None