fix: escape analysis — loop elements, (; v) shorthand, short-form nested defs

docs/src/safety/compile-time.md

@@ -67,10 +67,10 @@ in expression starting at myfile.jl:1
 
 ### Incidental Tail Patterns
 
-Beyond the direct-return patterns above, three additional tail shapes are
+Beyond the direct-return patterns above, four additional tail shapes are
 checked — they don't *look* like a `return`, but the scope's last expression
 still evaluates to a pool-backed array, which then escapes as the scope's
-value:
+value (also when wrapped in `identity(...)`):
 
 ```julia
 @with_pool pool begin
@@ -87,6 +87,10 @@ end
     v = acquire!(pool, Float64, 100)
     y = v                           # ← assignment tail
 end
+
+@with_pool pool begin
+    (acquire!(pool, Float64, 100), 1)   # ← literal holding a direct acquire call
+end
 ```
 
 Fix: end the block with `nothing` if the value is meant to be discarded —
@@ -145,3 +149,26 @@ end
 ```
 
 That's what the [runtime layer](runtime.md) is for.
+
+Loops and comprehensions bind *elements*, not the iterable. Reading an element of
+a pool-backed array is a plain value, so this is silent and correct:
+
+```julia
+@with_pool pool begin
+    order = acquire!(pool, Int, n)
+    contact = 0
+    for k in order          # `k` is an element (e.g. an Int), not pool memory
+        k == target && (contact = k)
+    end
+    (; contact)             # fine — no warning, no guard
+end
+```
+
+Iterating a container *literal* of pool arrays (`for c in (v, w)`) is tracked:
+`c` is a pool-backed array inside the loop. Iterating a container held in a
+*variable* (`t = (v, w); for c in t`) and destructuring iteration
+(`for (a, b) in ...`) are not tracked at compile time. With the default
+`runtime_check = 0`, nothing guards those shapes. With `runtime_check = 1`, the
+[runtime layer](runtime.md) inspects the scope's return value at exit and throws
+`PoolRuntimeEscapeError` if it still references pool memory — including through
+a `Tuple` or `NamedTuple` — on every supported Julia version.

ext/AdaptiveArrayPoolsCUDAExt/debug.jl

@@ -310,11 +310,11 @@ end
     )
     calls = [
         :(
-                _check_tp_cuda_overlap(
-                    getfield(pool, $(QuoteNode(f))), arr_ptr, arr_end,
-                    current_depth, pool, return_site, original_val
-                )
+            _check_tp_cuda_overlap(
+                getfield(pool, $(QuoteNode(f))), arr_ptr, arr_end,
+                current_depth, pool, return_site, original_val
             )
+        )
             for f in GPU_FIXED_SLOT_FIELDS
     ]
     return quote

ext/AdaptiveArrayPoolsMetalExt/debug.jl

@@ -310,11 +310,11 @@ end
     )
     calls = [
         :(
-                _check_tp_metal_overlap(
-                    getfield(pool, $(QuoteNode(f))), arr_buf, arr_off, arr_end,
-                    current_depth, pool, return_site, original_val
-                )
+            _check_tp_metal_overlap(
+                getfield(pool, $(QuoteNode(f))), arr_buf, arr_off, arr_end,
+                current_depth, pool, return_site, original_val
             )
+        )
             for f in METAL_FIXED_SLOT_FIELDS
     ]
     return quote

src/debug.jl

@@ -193,11 +193,11 @@ end
     )
     calls = [
         :(
-                _check_tp_pointer_overlap(
-                    getfield(pool, $(QuoteNode(f))), arr_ptr, arr_end,
-                    current_depth, pool, return_site, original_val
-                )
+            _check_tp_pointer_overlap(
+                getfield(pool, $(QuoteNode(f))), arr_ptr, arr_end,
+                current_depth, pool, return_site, original_val
             )
+        )
             for f in FIXED_SLOT_FIELDS
     ]
     return quote