render: time the submit-record loop and pre-size per-frame bookkeeping

src/skate3_native_scene.cpp

@@ -8236,6 +8236,9 @@ void BuildFrameScene(uint8_t* base, const SubmitRecord* records, size_t count) {
   FrameScene scene;
   scene.items.reserve(count);
   std::unordered_set<uint32_t> seen;
+  // Pre-size the per-frame bookkeeping: these fill with thousands of
+  // entries every frame, and growing from empty rehashes repeatedly.
+  seen.reserve(count);
   // Dynamic contexts are submitted several times per frame (once per pass);
   // each submission carries that pass's culled island list. Keep the fullest
   // one; a shadow-pass list can be missing body parts the main view needs.
@@ -8253,6 +8256,11 @@ void BuildFrameScene(uint8_t* base, const SubmitRecord* records, size_t count) {
     for (const DrawEntry& e : d.draws) n += e.index_count;
     return n;
   };
+  const bool wloop_prof = REXCVAR_GET(skate3_native_render_scene_perf_items);
+  const auto wloop_t0 = wloop_prof ? PerfClock::now() : PerfClock::time_point{};
+  if (wloop_prof) {
+    g_bi_records.fetch_add(count, std::memory_order_relaxed);
+  }
   for (size_t i = 0; i < count; ++i) {
     const SubmitRecord& r = records[i];
     // Primary opaque list of the chosen view only; other lists (shadow
@@ -8400,6 +8408,9 @@ void BuildFrameScene(uint8_t* base, const SubmitRecord* records, size_t count) {
       }
     }
   }
+  if (wloop_prof) {
+    g_pw_bi_wloop.Add(PerfNsSince(wloop_t0));
+  }
   // Cross-frame palette rescue + cache refresh (see g_bones_cache_ctx and
   // g_ropa_state_cache): published copies refresh their cache entries; a
   // refused/pending capture re-publishes with the cached palette (one frame

src/skate3_native_scene_gpu.cpp

@@ -7081,7 +7081,8 @@ void LogFrameStats(const FrameScene& scene, uint64_t frames, uint32_t drawn,
           "grid[valid={} age={}] "
           "stages_us[mesh={:.2f} tex={:.2f} const={:.2f} submit={:.2f} n={}] "
           "build_us[core n={} av={:.2f} fp n={} av={:.2f} walk n={} av={:.2f} "
-          "fetch av={:.2f}] retain[{:.2f}/{:.2f}ms app={} live={}]",
+          "fetch av={:.2f}] wloop[{:.2f}/{:.2f}ms recs={}] "
+          "retain[{:.2f}/{:.2f}ms app={} live={}]",
           g_pw_di_in.count.load(std::memory_order_relaxed), avg_us(g_pw_di_in),
           g_pw_di_in.MaxMs(),
           g_pw_di_occ.count.load(std::memory_order_relaxed),
@@ -7101,8 +7102,10 @@ void LogFrameStats(const FrameScene& scene, uint64_t frames, uint32_t drawn,
           avg_us(g_pw_bi_core),
           g_pw_bi_fp.count.load(std::memory_order_relaxed), avg_us(g_pw_bi_fp),
           g_pw_bi_walk.count.load(std::memory_order_relaxed),
-          avg_us(g_pw_bi_walk), avg_us(g_pw_bi_fetch), g_pw_bi_retain.AvgMs(),
-          g_pw_bi_retain.MaxMs(),
+          avg_us(g_pw_bi_walk), avg_us(g_pw_bi_fetch), g_pw_bi_wloop.AvgMs(),
+          g_pw_bi_wloop.MaxMs(),
+          g_bi_records.exchange(0, std::memory_order_relaxed),
+          g_pw_bi_retain.AvgMs(), g_pw_bi_retain.MaxMs(),
           g_retained_appended.exchange(0, std::memory_order_relaxed),
           g_retained_live.load(std::memory_order_relaxed));
     }
@@ -7114,7 +7117,8 @@ void LogFrameStats(const FrameScene& scene, uint64_t frames, uint32_t drawn,
                           &g_pw_di_in, &g_pw_di_occ, &g_pw_di_out,
                           &g_pw_di_mesh, &g_pw_di_tex, &g_pw_di_const,
                           &g_pw_di_submit, &g_pw_bi_core, &g_pw_bi_fp,
-                          &g_pw_bi_walk, &g_pw_bi_fetch, &g_pw_bi_retain}) {
+                          &g_pw_bi_walk, &g_pw_bi_fetch, &g_pw_bi_wloop,
+                          &g_pw_bi_retain}) {
       w->Reset();
     }
     // Refill the settle offender-log + slow-frame budgets for the next window.

src/skate3_native_scene_state.h

@@ -973,6 +973,13 @@ inline PerfWindow g_pw_bi_core;    // BuildItemFromMeshCached total (hit validat
 inline PerfWindow g_pw_bi_fp;      // ComputeItemFingerprint on the cache-hit path
 inline PerfWindow g_pw_bi_walk;    // full BuildItemFromMesh rebuilds (cache miss)
 inline PerfWindow g_pw_bi_fetch;   // AdoptDrawFetchOverrides per-frame guest reads
+// Whole submit-record loop of BuildFrameScene (one Add per frame). Contains
+// bi_core/bi_fetch plus everything they do not: per-record hash-set
+// bookkeeping, the dynamic merge, draw-list copies and DrawItem churn.
+// build - b2d - bspl - wloop - bpal - retain - remainder = the post-loop
+// tail (camera smoothing, capture blocks, publish).
+inline PerfWindow g_pw_bi_wloop;
+inline std::atomic<uint64_t> g_bi_records{0};  // submit records per window
 // Off-screen retention pass (always measured, one Add per frame: two clock
 // reads): its cost and how many items it re-appends beyond the guest's own
 // submission are core visibility-culling telemetry, not deep profiling.