1 files changed, 39 insertions, 12 deletions

diff --git a/src/engine/thread.c b/src/engine/thread.c
index 5891558..1bb9a83 100644
--- a/src/engine/thread.c
+++ b/src/engine/thread.c
@@ -40,7 +40,7 @@ void threads_go(struct threads *handler) {
 
 void *child_thread(void *_args) {
   struct child_thread_args *args = (struct child_thread_args*)_args;
-  args->stop = 0;
+  atomic_store(&args->stop, 0);
 
   atomic_bool stop = false;
   u64 nodes = 0;
@@ -71,7 +71,7 @@ void *child_thread(void *_args) {
     // iteration's block or the search is stopped
     if (sp == NULL) {
       static _Thread_local int spins = 0;
-      if (++spins == 5000000) fprintf(stderr, "CHILD %d STUCK stop=%d\n", args->id, atomic_load(&args->stop));
+      if (++spins == 5000000) fprintf(stderr, "CHILD %d STUCK stop=%d (ptr=%p)\n", args->id, atomic_load(&args->stop), (void*)&args->stop);
       sched_yield();
       continue;
     }
@@ -89,8 +89,14 @@ void *child_thread(void *_args) {
       position_t child = sp->position_snapshot;
       position_make_move(&child, move);
 
+      u64 prev_nodes = nodes;
       search_result_t res = search(child, sp->depth - 1, alpha, beta, &ctx);
 
+      // §2.1 thrashing guard: track how many nodes were searched at this
+      // split point so that split_point_worth_sharing() can refuse HELP
+      // on very shallow / barely-searched blocks.
+      atomic_fetch_add(&sp->nodes_searched, (int)(nodes - prev_nodes));
+
       // Share (§2.8): negamax — our value flips sign for the parent side
       split_point_share(sp, -res.score, move);
       if (-res.score > best_score) {
@@ -125,7 +131,7 @@ void *child_thread(void *_args) {
 // tree
 void *root_thread(void *_args) {
   struct root_thread_args *args = (struct root_thread_args*)_args;
-  args->stop = 0;
+  atomic_store(&args->stop, 0);
 
   atomic_bool stop = false;
   u64 nodes = 0;
@@ -149,6 +155,12 @@ void *root_thread(void *_args) {
   // iterative deepening: search at increasing depths, publishing an info
   // message per completed depth so the UCI thread can emit `info ... pv`.
   for (int depth = 1; depth <= max_depth; depth++) {
+    // propagate external stop signal into the search context so that
+    // search_should_stop() picks it up at the next node.
+    if (atomic_load(&args->stop)) {
+      atomic_store(&stop, true);
+      break;
+    }
     if (atomic_load(&stop)) break;
 
     // root move list (DTS §3.4: splitting at the root is always safe/PV)
@@ -156,7 +168,8 @@ void *root_thread(void *_args) {
     moves_init(&moves);
     get_legal_moves(&moves, *args->position);
 
-    bool can_split = moves.length > 1 && args->split_points != NULL;
+    bool can_split = moves.length > 1 && args->split_points != NULL
+      && depth >= 3; // §2.1 thrashing guard: don't split at depth < 3
     split_point_t *sp = can_split
       ? __pool_alloc(args->split_points, args->sp_pool_capacity)
       : NULL;
@@ -221,6 +234,7 @@ void *root_thread(void *_args) {
 
       result.score = atomic_load(&sp->best_score);
       result.best_move = sp->best_move;
+      moves_deinit(sp->moves);
       // the move list now belongs to the pool block (released on its next
       // reuse), so there is nothing to free here
     } else {
@@ -230,17 +244,28 @@ void *root_thread(void *_args) {
     }
 
     if (atomic_load(&stop)) break;
+    if (atomic_load(&args->stop)) {
+      atomic_store(&stop, true);
+      break;
+    }
 
     message->depth = depth;
     message->nodes = nodes > INT_MAX ? INT_MAX : (int)nodes;
     message->score_cp = result.score;
     message->mate = __score_to_mate(result.score);
-    // pv is already initialized by engine_message_init / handler_push.
     message->pv.count = 0;
-    add_ipc_move(&message->pv, (ipc_move_t) {
-      result.best_move.from,
-      result.best_move.to,
-    });
+    move_t pv_buf[32];
+    int pv_len = reconstruct_pv(*args->position, *args->ttable, pv_buf, 32);
+    if (pv_len > 0) {
+      for (int i = 0; i < pv_len; i++) {
+        add_ipc_move(&message->pv, (ipc_move_t) { pv_buf[i].from, pv_buf[i].to });
+      }
+    } else {
+      add_ipc_move(&message->pv, (ipc_move_t) {
+        result.best_move.from,
+        result.best_move.to,
+      });
+    }
 
     if (!should_continue(message, args->go_args)) {
       break;
@@ -257,8 +282,8 @@ void *root_thread(void *_args) {
   engine_message_mark_ready(message);
 
   // signal workers to stop their loops, then mark this thread as finished.
-  for (int i = 0; i < args->workers_count; i++) {
-    atomic_store(&args->workers_stop[i], 1);
+  for (int i = 1; i <= args->workers_count; i++) {
+    atomic_store(&args->handler->args[i].child_thread_args.stop, 1);
   }
   atomic_store(&args->stop, 1);
   atomic_store(&args->finished, 1);
@@ -291,6 +316,7 @@ static bool __root_thread_run(
 
   ttable_t tt;
   ttable_init(&tt, ttable_log);
+  args->ttable = &tt;
 
   root_thread(args);
 
@@ -350,7 +376,7 @@ bool test_root_thread_emits_info_and_bestmove() {
   // every published message must have a valid depth and on-board pv move.
   for (engine_message_t *m = messages; m != NULL; m = m->next) {
     ok &= m->depth >= 1 && m->depth <= 3;
-    ok &= m->pv.count == 1;
+    ok &= m->pv.count >= 1;
     ok &= m->pv.data[0].from >= 0 && m->pv.data[0].from < 64;
     ok &= m->pv.data[0].to >= 0 && m->pv.data[0].to < 64;
   }
@@ -694,6 +720,7 @@ static void __root_split_env_deinit(
     engine_message_handler_deinit(&env->message_handlers[i]);
   }
   ttable_deinit(env->handler.ttable);
+  threads_cleanup(&env->handler);
   threads_deinit(&env->handler);
 }