fix(gc): deserializer composite-rebuild rooting hole + TAP pin follow-through

Smell-audited: 2: the gc_torture failure A/B-bisected to the D-563(b)
commit, but the ROOT CAUSE is a pre-existing latent hole my change was
first to mass-exercise: serialize.readValue's composite arms (list /
vector / array_map / hash_set) grow `out` (+ in-flight key / cons buf)
as Zig LOCALS across per-element readValue/assoc/conj allocs — an
alloc-torture collect in the window sweeps the in-progress structure
(recycled ArrayMap count → the "index 16, len 16" trace walk panic).
Per-def meta constants (D-563b) made every lazy-ns bytecode replay
deserialize composites, exposing it at 6/10 under
CLJW_GC_TORTURE_ALLOC=1. Fix = the established fabrication-region
bracket (ADR-0150) on all four arms: 0/12 post-fix, full gc_torture
suite green. Composite constants are small — deferring the collect to
the bracket end is bounded.

Also: phase15_clojure_test_tap pins updated for the D-563(b)
` (file:line)` suffix (clj's TAP emits it too — oracle-checked;
deftest-var-sourced per the narrowed AD-041).

Author: chaploud

src/eval/bytecode/serialize.zig

@@ -551,6 +551,13 @@ fn readValue(ctx: *ReadCtx, r: *ByteReader) DeserializeError!Value {
             // primitive at runtime.
             const buf = std.heap.page_allocator.alloc(Value, n) catch return DeserializeError.OutOfMemory;
             defer std.heap.page_allocator.free(buf);
+            // Fabrication region: `buf` elements + the growing `lst` are Zig
+            // locals across the nested readValue/cons allocs — a collect in
+            // the window would sweep the in-progress structure (the D-563b
+            // json_nested alloc-torture crash class). Composite constants are
+            // small; deferring the collect to the bracket end is safe.
+            rt.gc.enterFabrication();
+            defer rt.gc.exitFabrication();
             var i: u32 = 0;
             while (i < n) : (i += 1) buf[i] = try readValue(ctx, r);
             var lst: Value = .nil_val;
@@ -563,6 +570,10 @@ fn readValue(ctx: *ReadCtx, r: *ByteReader) DeserializeError!Value {
         },
         .vector => {
             const n = try r.readU32();
+            // Fabrication region: see the .list arm — `out` is unrooted
+            // across the per-element readValue/conj allocs.
+            rt.gc.enterFabrication();
+            defer rt.gc.exitFabrication();
             var out = vector_mod.empty();
             var i: u32 = 0;
             while (i < n) : (i += 1) {
@@ -573,6 +584,10 @@ fn readValue(ctx: *ReadCtx, r: *ByteReader) DeserializeError!Value {
         },
         .array_map => {
             const n = try r.readU32();
+            // Fabrication region: see the .list arm — `out` + the in-flight
+            // k are unrooted across the per-entry readValue/assoc allocs.
+            rt.gc.enterFabrication();
+            defer rt.gc.exitFabrication();
             var out = map_mod.empty();
             var i: u32 = 0;
             while (i < n) : (i += 1) {
@@ -584,6 +599,9 @@ fn readValue(ctx: *ReadCtx, r: *ByteReader) DeserializeError!Value {
         },
         .hash_set => {
             const n = try r.readU32();
+            // Fabrication region: see the .list arm.
+            rt.gc.enterFabrication();
+            defer rt.gc.exitFabrication();
             var out = set_mod.empty();
             var i: u32 = 0;
             while (i < n) : (i += 1) {

test/e2e/phase15_clojure_test_tap.sh

@@ -32,18 +32,18 @@ EOF
 # run-tests emits the full lifecycle event set — :begin-test-ns, :begin/:end-test-var
 # (per test), :end-test-ns — which tap's `:default` renders as `# {…}` comment
 # lines, EXACTLY as clj's tap does (clj-verified 2026-06-21). cljw omits only the
-# ` (file:line)` source suffix on ok/not-ok and the ns identity-hash in the
-# begin/end-test-ns map (no source location / JVM identity — AD-041 / AD-002);
-# every other byte matches clj. Names render via testing-vars-str `(test-name)`;
+# the ns identity-hash in the begin/end-test-ns map (no JVM identity —
+# AD-002); the ok/not-ok ` (file:line)` suffix is PRESENT since D-563(b)
+# (deftest-var-sourced; clj's is assertion-stack-sourced — narrowed AD-041). Names render via testing-vars-str `(test-name)`;
 # `=` actual is the bare (not (= …)) form (D-463).
 want='# {:type :begin-test-ns, :ns user}
 # {:type :begin-test-var, :var #'\''user/pass-t}
-ok (pass-t)
+ok (pass-t) (<stdin>:2)
 # expected:(= 1 1)
 #   actual:(= 1 1)
 # {:type :end-test-var, :var #'\''user/pass-t}
 # {:type :begin-test-var, :var #'\''user/fail-t}
-not ok (fail-t)
+not ok (fail-t) (<stdin>:3)
 # expected:(= 1 2)
 #   actual:(not (= 1 2))
 # {:type :end-test-var, :var #'\''user/fail-t}