fix(eval): let the strong-rabbit bonus grow with support

rabbit_strength() scales a supported rabbit's bonus by power /
full_strong, both ints, so the factor was 0 unless power had been
clamped to full_strong, and then 1: a step instead of a ramp. Divide as
reals. On 20,000 random positions the strong branch ran 11,658 times
and gave 0 in 3,137 of them where the intended factor is well above 0.

Also fix the rabbit square table's third row, which had 0.13 on c3
where its mirror f3 has 0.14, so the term changed under a file or
colour flip (up to 7 points, in 312 of 7093 random positions). Now no
random position's rabbit strength changes under either flip.

dub test has two new eval_tests: a rabbit on c3 and its mirror on f3
score the same, and a rabbit on c6 gets more bonus with a cat on c5
than on c4, and more than 0 (both fail on the old code). Over the
602,518 game positions of a 6000-game test the rabbit term changed in
38.8%, by a median of 21 points (90th percentile 137, max 1086), evenly
between the colours.

Test: 4452 games against f68264a at 1s/30s/100/0/100t, hash 64:
+5.5 Elo [-4.9, +15.9] (95%), SPRT [0, 10] undecided (LLR 0.16),
stopped once it had under a 5% chance of deciding within 3000 pairs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Author: Janzert

CHANGELOG.md

@@ -7,6 +7,17 @@ drawn from the commit history.
 
 ## Unreleased
 
+- **Fixed:** the bonus for a strong advanced rabbit was all or nothing.
+  Its share of support was an integer division, so it was 0 unless the
+  rabbit had full support, and then 1; it now grows with the support as
+  intended. The rabbit square table also gave c3 (and silver's f6) 0.13
+  where its mirror has 0.14, so the term changed when the board was
+  mirrored; it no longer does. Over the game positions of a 6000-game
+  test the rabbit term changed in 38.8%, by a median of 21 points (90th
+  percentile 137). Test: 4452 games against the build before, at
+  1s/move and a 100-turn limit: +5.5 Elo [−4.9, +15.9] (95%), SPRT
+  [0, 10] undecided (LLR 0.16), stopped once it had under a 5% chance
+  of deciding within 3000 pairs.
 - **Fixed:** a rare crash while reporting the principal variation
   during a search (the default multithreaded engine). The report walks
   the hash table while the search thread writes it, and read each

source/staticeval.d

@@ -441,7 +441,7 @@ class StaticEval
         // Try to encourage rabbits to the edges.
         static immutable real[] rsq = [0.0L, 0, 0, 0, 0, 0, 0, 0,
               0.05L, 0.07L, 0.09L, 0.10L, 0.10L, 0.09L, 0.07L, 0.05L,
-              0.11L, 0.12L, 0.14L, 0.15L, 0.15L, 0.13L, 0.12L, 0.11L,
+              0.11L, 0.12L, 0.14L, 0.15L, 0.15L, 0.14L, 0.12L, 0.11L,
               0.10L, 0.11L, 0.13L, 0.14L, 0.14L, 0.13L, 0.11L, 0.10L,
               0.09L, 0.10L, 0.11L, 0.12L, 0.12L, 0.11L, 0.10L, 0.09L,
               0.09L, 0.10L, 0.10L, 0.11L, 0.11L, 0.10L, 0.10L, 0.09L,
@@ -534,7 +534,7 @@ class StaticEval
                 {
                     int full_strong = STRONG_SPAN + power_balance;
                     power = (power < full_strong) ? power : full_strong;
-                    real sfactor = power / full_strong;
+                    real sfactor = cast(real) power / full_strong;
                     real rv = rankval[s][rix/8];
                     real rval = STRONG_VAL * rv * sfactor;
                     if (rbit & TRAPS)

tests/eval_test.d

@@ -79,3 +79,35 @@ unittest
     assert(score > 0);
     assert(score == capture_threat(mirror));
 }
+
+@("rabbit strength doesn't depend on file order")
+unittest
+{
+    // A gold rabbit on c3 and its mirror on f3.
+    enum board = "[e       " ~ "        " ~ "        " ~ "        "
+        ~ "        " ~ "  R     " ~ "        " ~ "        ]";
+    enum mirror = "[       e" ~ "        " ~ "        " ~ "        "
+        ~ "        " ~ "     R  " ~ "        " ~ "        ]";
+    assert(evaluator(board).rabbit_strength()
+            == evaluator(mirror).rabbit_strength());
+}
+
+@("strong rabbit bonus grows with its support")
+unittest
+{
+    // A gold rabbit on c6 with a gold cat behind it, close (c5) or one
+    // square further (c4). Neither gives full support, so both get part
+    // of the bonus, the closer cat more.
+    enum close = "[        " ~ "        " ~ "  R     " ~ "  C     "
+        ~ "        " ~ "        " ~ "        " ~ "       e]";
+    enum far = "[        " ~ "        " ~ "  R     " ~ "        "
+        ~ "  C     " ~ "        " ~ "        " ~ "       e]";
+    StaticEval ev_close = evaluator(close);
+    StaticEval ev_far = evaluator(far);
+    ev_close.rweak_w = 0;
+    ev_far.rweak_w = 0;
+    real s_close = ev_close.rabbit_strength();
+    real s_far = ev_far.rabbit_strength();
+    assert(s_far > 0);
+    assert(s_close > s_far);
+}