Land super-lazy exit, material-based lazy floor, qsearch futility rework

Brings in the last remaining piece from stash@{0}: the super-lazy exit
point (material-only pre-check before the regular lazy gate), a
material-bucket floor under the regular lazy exit's margin
(iSwingFloorByArmy), and search.c's qsearch futility rework
(FUTILITY_BASE_MARGIN_BY_SOURCE, indexed by which Eval() exit tier
produced the score). Required Eval()'s signature change from a single
SCORE* to SCORE(*)[2] (positional estimate per side instead of one
munged magnitude) -- search.c's futility margin folds in
rgiPositional[pos->uToMove], which the earlier bad-trades investigation
found to be a meaningfully predictive signal.

search.c and chess.h brought in wholesale from the stash (both were
either completely untouched by prior commits or contained no
divergence worth preserving). eval.c required hand-merging on top of
this session's already-applied bad-trades fix, B-over-N removal, and
xColor/reorder cleanups -- ported the super-lazy exit block, the
regular-lazy material floor, the per-color piPositional writes (all
three exit sites: super-lazy, regular-lazy x2, full-eval), the
super-lazy calibration harness (RecordSuperLazyMarginSafetySwing,
dual-regime DumpMarginSafetyCalibration), and moved uArmyScaler/
uNumTrapped initialization to match the new ordering the super-lazy
exit depends on.

Verified: clean release + DEBUG build (only the previously-flagged
_EvalTrappedPieces warning), DEBUG smoke test pass, and a 40-game
st1 match against clean 434fa04 (score 0.487, llr -0.04) -- landing
this margin machinery as-is from the stash, before any retuning, does
not regress strength on its own. This confirms the original
regression (0.15-0.225 score seen early in this investigation) was
fully explained by the bad-trades unsigned-underflow bug, not by these
margins being unsound.

Values are the original, as-derived-from-calibration ones (see
inline comments: SUPER_LAZY_MARGIN_BY_ARMY from 100 positions/sd8
with ~15-20% headroom, iSwingFloorByArmy from 1500 positions/sd8
with ~25% headroom, FUTILITY_BASE_MARGIN_BY_SOURCE from a separate
1500-position/sd6 surprise-rate calibration). Not yet retuned --
suspected to carry more headroom than necessary, which costs real
search speed (speed=depth). Next step: re-run the margin-safety
calibration fresh against the current build and trim the super-lazy
and regular-lazy floor headroom down from the built-in database,
leaving the qsearch futility margins alone (different calibration
method, already risk-tolerant by construction).

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Ka9o3S2eKqh4jNfmxVZ6fH

1 files changed, 111 insertions, 50 deletions

diff --git a/src/chess.h b/src/chess.h
index 702defb..78c0f94 100755
--- a/src/chess.h
+++ b/src/chess.h
@@ -941,19 +941,9 @@ typedef struct _COUNTERS
         UINT64 u64AvoidNullSuccess;
         UINT64 u64AvoidNullFailures;
 #endif
-        UINT64 u64EvalHashHits;
         UINT64 u64LazyEvals;
-        UINT64 u64FullEvals;
-        // Placeholder counters for root.c's per-tier eval-exit
-        // reporting -- eval.c's super-lazy exit itself hasn't been
-        // re-applied yet (see stash), so these always read 0 for now;
-        // that's accurate, not a stub bug, since no super-lazy exit
-        // exists in this build to increment them.
         UINT64 u64SuperLazyEvals;
-        UINT64 u64CyclesSuperLazyExit;
-        UINT64 u64CyclesLazyExit;
-        UINT64 u64CyclesFullEvalExit;
-        UINT64 u64CyclesEvalSuperLazy;
+        UINT64 u64FullEvals;
         UINT64 u64CyclesInEval;
 
         //
@@ -1021,6 +1011,30 @@ typedef struct _COUNTERS
         UINT64 u64CyclesEvalLazyDecision;
         UINT64 u64CyclesEvalCountKingSafetyDefects;
         UINT64 u64CyclesEvalFileStormDefects;
+
+        //
+        // 2026-09-07: the super-lazy exit (eval.c, before the regular
+        // LAZY_EVAL block) is a much cheaper material-only check than
+        // the regular lazy decision, so it gets its own cycle counter
+        // rather than folding into u64CyclesEvalLazyDecision -- a hit
+        // here never reaches the regular lazy-decision code at all.
+        //
+        UINT64 u64CyclesEvalSuperLazy;
+
+        // Per-exit-path *total call cost*, i.e. the full entry-to-exit
+        // elapsed time for a call that left via that path -- not to be
+        // confused with the sub-timers above, which measure a single
+        // segment's cost regardless of how the call eventually exited.
+        // These three are mutually exclusive by construction (each
+        // Eval() call adds to exactly one) and sum to u64CyclesInEval,
+        // so u64CyclesInEval / (u64SuperLazyEvals + u64LazyEvals +
+        // u64FullEvals) is a real overall average, and each bucket
+        // divided by its own matching count (u64SuperLazyEvals,
+        // u64LazyEvals, u64FullEvals) gives a clean per-path average --
+        // see root.c's eval-exit-breakdown report.
+        UINT64 u64CyclesSuperLazyExit;
+        UINT64 u64CyclesLazyExit;
+        UINT64 u64CyclesFullEvalExit;
     }
     tree;
 
@@ -1241,6 +1255,24 @@ typedef struct _SEARCHER_THREAD_CONTEXT
     // called on.
     FLAG fCalibrateCandidate;
 #endif
+    // Which Eval() exit tier supplied the piPositional value most
+    // recently returned to this thread -- set inside Eval() itself
+    // (all three exit points), read right after the Eval() call in
+    // QSearch. Was CALIBRATE_QSEARCH_FUTILITY-only; promoted to
+    // unconditional 2026-09-08 since QSearch's real futility margin
+    // now indexes FUTILITY_BASE_MARGIN_BY_SOURCE with it, not just
+    // the calibration harness. See EVAL_POSITIONAL_SOURCE_* below.
+    ULONG uLastPositionalSource;
+#ifdef CALIBRATE_QSEARCH_FUTILITY
+    // TRUE for the duration of a diagnostic "what if we hadn't pruned
+    // this move" re-search -- search.c's qsearch futility gate checks
+    // this and forces iFutility wide open so the diagnostic subtree
+    // itself isn't contaminated by the same pruning being measured.
+    // Never set recursively: a rejection can't occur while this is
+    // already TRUE (iFutility being wide open means nothing gets
+    // rejected), so there's no risk of runaway nested diagnosis.
+    FLAG fDiagUnprunedSubtree;
+#endif
     CHAR szLastPV[SMALL_STRING_LEN_CHAR];
 }
 SEARCHER_THREAD_CONTEXT;
@@ -2820,14 +2852,27 @@ IsDraw(SEARCHER_THREAD_CONTEXT *ctx);
 // search.c
 //
 #define QPLIES_OF_NON_CAPTURE_CHECKS (2)
-#define FUTILITY_BASE_MARGIN         (50)
-// Compatibility aliases for root.c/main.c's per-tier reporting, which
-// expects these three names -- search.c hasn't been split into
-// per-tier margins yet (still one flat FUTILITY_BASE_MARGIN), so all
-// three alias the same value until that split is re-applied.
-#define FUTILITY_BASE_MARGIN_FULL      FUTILITY_BASE_MARGIN
-#define FUTILITY_BASE_MARGIN_LAZY      FUTILITY_BASE_MARGIN
-#define FUTILITY_BASE_MARGIN_SUPERLAZY FUTILITY_BASE_MARGIN
+
+// Which Eval() exit tier supplied a given piPositional value --
+// SEARCHER_THREAD_CONTEXT.uLastPositionalSource above. Was diagnostic-
+// only (CALIBRATE_QSEARCH_FUTILITY); promoted to unconditional
+// 2026-09-08 once FUTILITY_BASE_MARGIN_BY_SOURCE (search.c) made it
+// load-bearing for real search behavior, not just measurement.
+#define EVAL_POSITIONAL_SOURCE_FULL      (0)
+#define EVAL_POSITIONAL_SOURCE_LAZY      (1)
+#define EVAL_POSITIONAL_SOURCE_SUPERLAZY (2)
+#define EVAL_POSITIONAL_SOURCE_COUNT     (3)
+
+// Per-tier qsearch futility base margins -- see search.c's
+// FUTILITY_BASE_MARGIN_BY_SOURCE for the full derivation. Named here
+// (rather than only as array entries in search.c) so eval.c's
+// piPositional floor -- which only ever applies on the full-eval
+// exit -- can reference the matching tier's constant directly instead
+// of a stale flat value.
+#define FUTILITY_BASE_MARGIN_FULL      (450)
+#define FUTILITY_BASE_MARGIN_LAZY      (225)
+#define FUTILITY_BASE_MARGIN_SUPERLAZY (275)
+
 // Measured: disabling this entirely (see lmr_testing/RESULTS.md) is a
 // clear net loss across ringers/confident_quick/hard_quick, so IID itself
 // is load-bearing. The "is the top move crappy" gate in search.c's DO_IID
@@ -3144,7 +3189,7 @@ PawnHashLookup(SEARCHER_THREAD_CONTEXT *ctx);
 //
 #define LAZY_EVAL
 #define LAZE_EVAL_BASE_SCORE 10
-#define LAZY_EVAL_BASE_MARGIN (75) // cheap material-only lazy exit margin;
+#define LAZY_EVAL_BASE_MARGIN (75)  // cheap material-only lazy exit margin;
                                     // widened by EstimatePositionalScore
                                     // if this isn't enough on its own
 // board_representation/EVAL.md section 9 (2026-09-06): below this
@@ -3162,13 +3207,6 @@ PawnHashLookup(SEARCHER_THREAD_CONTEXT *ctx);
 extern const int g_iAhead[2];
 extern const int g_iBehind[2];
 
-// No-op placeholder: eval.c's ROOK_FULL_HALF_OPEN_BONUS static cache
-// (the thing InitEval() is meant to (re)build, called at startup and
-// after every DNA reload) hasn't been re-applied yet -- see stash.
-// Nothing to initialize until that cache exists.
-void
-InitEval(void);
-
 ULONG
 DNABufferSizeBytes();
 
@@ -3184,14 +3222,15 @@ ImportEvalDNA(char *p);
 FLAG
 ReadEvalDNA(char *szFilename);
 
-
+void
+InitEval();
 
 SCORE
-Eval(SEARCHER_THREAD_CONTEXT *, SCORE, SCORE, SCORE *);
+Eval(SEARCHER_THREAD_CONTEXT *, SCORE, SCORE, SCORE (*)[2]);
 
 FLAG
 _EvalPasserRacesAgainstLoneKings(POSITION *,
-                PAWN_HASH_ENTRY *);
+                                 PAWN_HASH_ENTRY *);
 
 ULONG
 CountKingSafetyDefects(POSITION *pos,
@@ -3212,6 +3251,22 @@ void
 DumpMarginSafetyCalibration(void);
 #endif
 
+#ifdef CALIBRATE_QSEARCH_FUTILITY
+// Which futility gate in _ShouldWeConsiderThisMove rejected a move --
+// see search.c's _QFutRecordSample. The recapture gate (used to be
+// QFUT_GATE_RECAPTURE) was retired 2026-09-08 along with the +100
+// recapture bonus itself -- see the "recapture-shaped" comment in
+// _ShouldWeConsiderThisMove for why (fRecaptureShaped never checked
+// mv.cTo == mvLast.cTo, so it wasn't testing real recaptures).
+#define QFUT_GATE_GENERIC_CAPTURE (0)
+#define QFUT_GATE_CHECK_ROOK      (1)
+#define QFUT_GATE_CHECK_BISHOP    (2)
+#define QFUT_GATE_COUNT           (3)
+
+void
+DumpQSearchFutilityCalibration(void);
+#endif
+
 //
 // testeval.c
 //
@@ -3352,22 +3407,13 @@ ParallelCompareUlong(ULONG uComparand, void *pComparators);
 ULONG CDECL
 ParallelCompareVector(void *pComparand, void *pComparators);
 
-void CDECL
-GetAttacks(SEE_LIST *pList,
-           POSITION *pos,
-           COOR cSquare,
-           ULONG uSide);
-
-void CDECL
-SlowGetAttacks(SEE_LIST *pList,
-               POSITION *pos,
-               COOR cSquare,
-               ULONG uSide);
-
-// board_representation/MIGRATION.md section 3: bbPieces/bbPawns-backed
-// GetAttacks primitive, verified correct (20,000-position sweep) and
-// faster than asm GetAttacks (0.53-0.89x cycles/call across
-// opening/middlegame/endgame -- see section 3's benchmark writeup).
+// board_representation/MIGRATION.md sections 3/6/7: bbPieces/bbPawns-
+// backed GetAttacks primitive. Verified correct (20,000-position
+// sweep) and faster than the old mailbox implementation (0.53-0.89x
+// cycles/call across opening/middlegame/endgame). Now the only
+// GetAttacks implementation -- the old asm/CROUTINES mailbox versions
+// (SlowGetAttacks, asm GetAttacks) were retired 2026-09-06 once all of
+// section 7's retirement criteria cleared.
 void CDECL
 _GetAttacksBB(SEE_LIST *pList,
               POSITION *pos,
@@ -3386,10 +3432,9 @@ _WhoAttacksSquareBB(POSITION *pos,
                      BITBOARD bbOccupied);
 
 // _GetAttacksBB is the only implementation as of 2026-09-06 (see
-// MIGRATION.md section 6) -- verified correct and faster than the old
-// asm/CROUTINES mailbox versions (SlowGetAttacks, asm GetAttacks),
-// which have been retired. Every call site written against the name
-// "GetAttacks" didn't need touching when the default changed.
+// MIGRATION.md section 7) -- this macro exists only so call sites
+// written against the name "GetAttacks" didn't need touching when the
+// old asm/CROUTINES mailbox versions were retired.
 #define GetAttacks _GetAttacksBB
 
 #ifdef _X86_
@@ -3596,6 +3641,12 @@ COMMAND(BenchCommand);
 void
 TestDraw(void);
 
+void
+TestRecogn(void);
+
+void
+TestRecognExhaustiveKNKP(void);
+
 //
 // probe.c
 //
@@ -3683,6 +3734,16 @@ RecognLookup(SEARCHER_THREAD_CONTEXT *ctx,
              SCORE *piScore,
              FLAG fProbeEGTB);
 
+//
+// Unregistered/experimental recognizers, exposed only so testrecogn.c
+// can exhaustively validate a candidate before it's ever re-registered
+// in InitializeInteriorNodeRecognizers. Not reachable from the normal
+// RecognLookup dispatch path.
+//
+ULONG
+_RecognizeKNKP(SEARCHER_THREAD_CONTEXT *ctx,
+               SCORE *piScore);
+
 SCORE
 GetRoughEvalScore(IN SEARCHER_THREAD_CONTEXT *ctx,
                   IN SCORE iAlpha,