diff --git a/docs/tickets.md b/docs/tickets.md
index e227841..6f7ef1d 100644
--- a/docs/tickets.md
+++ b/docs/tickets.md
@@ -689,3 +689,27 @@ symbolized stack; fix then.
 (profiler overlay/history strings are the heaviest string traffic in the
 crashing runs), or an engine-side heap read overrun in a small shared
 helper. Audit report: `shooter/docs/audit-round2.md` finding F1.
+
+## EN-023 — GI software path: colored bounce is unreachable 🟡
+
+**Why.** Round-2 audit (F4): on adapters without EXPERIMENTAL_RAY_QUERY
+(the dev box's Radeon 760M reports none — see the new boot log), the
+probe trace runs the SDF path, where the mesh-card lookup compares
+world-space hits against OBJECT-space AABBs (ssgi.rs broad phase) — every
+transformed instance falls through to the flat gray 0.55 analytic — and
+the software WSRC bake is analytic sun+sky with no geometry at all.
+Measured in-game: SSGI on/off is ≤2.5% achromatic luma, hue ratio
+unchanged to 4 decimals. The ~267 gi_only proxies feed a pipeline whose
+colored output cannot reach the screen on SW adapters, at ~1.6 ms GPU in
+combat.
+
+**Sketch.** (1) Transform SDF hits into instance object space (or
+instance AABBs into world space) in the card broad-phase so textured/
+tinted cards resolve. (2) Make the SW WSRC bake sample cards (or at
+least sun-shadowed ground/albedo bounce) instead of pure analytic.
+(3) Re-run the audit's GI A/B: acceptance = G/R hue shift at the
+building base and ≥8-10% luma in shaded receivers.
+
+**Interim option** (decision D2-B in the shooter audit): boot-time
+`setSsgiEnabled(false)` on SW adapters banks the ~1.6 ms until this
+lands; needs a backend query FFI or the game reading the new boot log.
diff --git a/native/shared/src/renderer/scene_pass.rs b/native/shared/src/renderer/scene_pass.rs
index 29fabf7..b24c59e 100644
--- a/native/shared/src/renderer/scene_pass.rs
+++ b/native/shared/src/renderer/scene_pass.rs
@@ -464,6 +464,15 @@ impl Renderer {
         if let Some(tid) = scene_color_tid {
             self.transient_pool.release(tid);
         }
+        // Round-2 audit: the depth snapshot was acquired every frame but
+        // NEVER released — the pool (which has no auto-reclaim by design)
+        // allocated a brand-new render-res Depth32Float every frame with
+        // refractive water on screen. ~8 MB/frame of texture creation plus
+        // an ever-growing slot scan; measured as translucent_pass CPU
+        // climbing 0.4 ms → 6.7 ms over ~50 s of play, in every session.
+        if let Some(tid) = scene_depth_tid {
+            self.transient_pool.release(tid);
+        }
         profiler.end("translucent_pass");
     }
     }
diff --git a/native/shared/src/renderer/transient.rs b/native/shared/src/renderer/transient.rs
index c5e6a73..dd7b8fb 100644
--- a/native/shared/src/renderer/transient.rs
+++ b/native/shared/src/renderer/transient.rs
@@ -187,6 +187,21 @@ impl TransientPool {
         }
 
         // Nothing matched — allocate a new slot.
+        //
+        // Leak watchdog: a steady-state frame needs a handful of slots.
+        // A count that keeps growing means some caller acquires without
+        // releasing (the pool has no auto-reclaim by design) — exactly
+        // the depth-snapshot leak the round-2 audit spent a day tracing
+        // from the profiler side. Warn at every power of two past 64 so
+        // the log stays quiet but the failure mode is named.
+        let n = self.slots.len();
+        if n >= 64 && n.is_power_of_two() {
+            eprintln!(
+                "bloom transient pool: {} slots allocated and growing — \
+                 an acquire without a matching release is leaking textures",
+                n
+            );
+        }
         let texture = device.create_texture(&wgpu::TextureDescriptor {
             label: Some("transient"),
             size: wgpu::Extent3d {