phase3: generalize trust layer for C/C++; wire cJSON target

Trust-layer generalization (zero-regression verified on tokei):
- TargetSpec [build] uses an explicit {out} isolation-dir placeholder
  instead of injected CARGO_TARGET_DIR, plus optional sanitizer/
  sanitizer_binary templates. Nothing assumes cargo anymore.
- verdict rebuild_pristine_baseline, sanitizer_check, and harnessd
  run_verdict candidate build all use {out} substitution. A target with
  no sanitizer build caps accepts at needs-human-review.
- tokei/comrak target.toml migrated to the {out} form (CARGO_TARGET_DIR=
  {out} + nightly ASan sanitizer command).

cJSON C target (v1.7.19, MIT):
- harness.c bench/golden driver (trust-layer; agent can patch workspace
  cJSON.c but not the rig), compiled with clang against the patched tree.
- 5.6 MB deterministic corpus (gen-corpus.sh) + MANIFEST/GOLDEN pins.
- gates = hermetic Unity/CTest suite (19 tests pass); sanitizer =
  ASan+UBSan; scripts/diff-fuzz-cjson.py for differential fuzzing.

coz causal profiler integration (SPEC §12):
- harness.c optional -DHOTPATH_COZ throughput progress point.
- scripts/install-coz.sh, coz-profile.sh, coz-summary.py; just coz/install-coz.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EcKPWQ3MQD4NdEHwar96u2

targets/cjson/harness.c

@@ -0,0 +1,73 @@
+/* hotpath bench/golden driver for the cjson target.
+ *
+ * TRUST-LAYER CODE — lives outside targets/cjson/workspace, so the agent's
+ * propose_patch allowlist (workspace-relative paths only) cannot touch it.
+ * It is compiled together with the (patched) workspace cJSON.c, so a patch
+ * changes the parser/printer under test but never this measurement rig.
+ *
+ * Workload: parse a JSON document `iters` times, serialize it unformatted
+ * each time, and fold every serialized byte into a rolling checksum so the
+ * optimizer cannot elide the work. Deterministic: the last serialization
+ * plus the checksum are written to stdout, giving golden-replay a
+ * byte-exact signal over both the parse AND the print path.
+ *
+ *   cjson-bench <file.json> [iters=40]
+ */
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include "cJSON.h"
+
+/* coz causal profiling (SPEC §12): compile with -DHOTPATH_COZ and link
+ * libcoz to mark each parse+print iteration as a throughput progress
+ * point, so `coz run` can measure which source lines, if sped up, would
+ * actually raise end-to-end throughput. A no-op in normal builds. */
+#ifdef HOTPATH_COZ
+#include <coz.h>
+#define HOTPATH_PROGRESS() COZ_PROGRESS
+#else
+#define HOTPATH_PROGRESS() ((void)0)
+#endif
+
+int main(int argc, char **argv) {
+    if (argc < 2) {
+        fprintf(stderr, "usage: %s <file.json> [iters]\n", argv[0]);
+        return 2;
+    }
+    long iters = (argc >= 3) ? atol(argv[2]) : 40;
+    if (iters < 1) iters = 1;
+
+    FILE *f = fopen(argv[1], "rb");
+    if (!f) { perror("fopen"); return 2; }
+    fseek(f, 0, SEEK_END);
+    long n = ftell(f);
+    fseek(f, 0, SEEK_SET);
+    if (n < 0) { fclose(f); return 2; }
+    char *buf = (char *)malloc((size_t)n + 1);
+    if (!buf) { fclose(f); return 2; }
+    if (fread(buf, 1, (size_t)n, f) != (size_t)n) { free(buf); fclose(f); return 2; }
+    buf[n] = '\0';
+    fclose(f);
+
+    unsigned long acc = 1469598103934665603UL; /* FNV-ish rolling checksum */
+    char *last = NULL;
+    for (long i = 0; i < iters; i++) {
+        cJSON *root = cJSON_Parse(buf);
+        if (!root) { fprintf(stderr, "parse failed\n"); free(buf); return 3; }
+        char *out = cJSON_PrintUnformatted(root);
+        if (!out) { cJSON_Delete(root); free(buf); return 3; }
+        for (char *p = out; *p; p++) acc = (acc ^ (unsigned char)*p) * 1099511628211UL;
+        if (i == iters - 1) last = out;
+        else free(out);
+        cJSON_Delete(root);
+        HOTPATH_PROGRESS(); /* one parse+print = one throughput unit (coz) */
+    }
+
+    if (last) {
+        fwrite(last, 1, strlen(last), stdout);
+        free(last);
+    }
+    printf("\nchecksum=%lu\n", acc);
+    free(buf);
+    return 0;
+}

targets/cjson/target.toml

@@ -0,0 +1,44 @@
+# cJSON — ultralightweight JSON parser/printer (C)
+# Selection vs SPEC §6: CPU-bound parse+serialize; hermetic Unity/CTest
+# suite vendored in-tree; deterministic scripted workload; MIT license;
+# active but deliberately "simple", no dedicated perf team — the richer
+# legacy-C hunting ground Phase 3 calls for.
+
+[source]
+repo = "https://github.com/DaveGamble/cJSON.git"
+commit = "c859b25da02955fef659d658b8f324b5cde87be3"   # v1.7.19
+license = "MIT"
+
+[fetch]
+# targets/*/workspace is gitignored; reproduce it with:
+#   git clone https://github.com/DaveGamble/cJSON.git workspace
+#   git -C workspace checkout v1.7.19
+# The bench/golden driver targets/cjson/harness.c and the Unity suite in
+# workspace/tests/ (unity vendored in-tree) are hermetic — no submodules.
+
+[build]
+# Compile the trust-layer harness against the (patched) workspace cJSON.c.
+# {out} is the isolation dir; the agent may patch only workspace/*, never
+# harness.c, so the measurement rig is fixed across baseline/candidate.
+baseline = "mkdir -p {out} && clang -O2 -DNDEBUG -I targets/cjson/workspace -o {out}/cjson-bench targets/cjson/harness.c targets/cjson/workspace/cJSON.c"
+binary = "{out}/cjson-bench"
+# ASan + UBSan (C/C++ analogue of the Rust nightly ASan path); run on
+# every would-be accept (SPEC §8), halting on any leak/UB.
+sanitizer = "mkdir -p {out} && clang -O1 -g -fsanitize=address,undefined -fno-sanitize-recover=all -I targets/cjson/workspace -o {out}/cjson-asan targets/cjson/harness.c targets/cjson/workspace/cJSON.c"
+sanitizer_binary = "{out}/cjson-asan"
+
+[bench]
+command = "{binary} corpora/cjson/input/big.json 2 > /dev/null"
+workload = "5.6 MB synthetic JSON (20k records: nested objects/arrays, escaped/unicode strings, ints, dense floats), parsed+serialized 2x, single thread"
+
+[gates]
+# Upstream Unity suite via CTest — parse_number/string/array/object,
+# print_*, misc_tests, compare/minify. Tests the PATCHED cJSON.c.
+tests = "cmake -S targets/cjson/workspace -B targets/cjson/testbuild -DENABLE_CJSON_TEST=On -DCMAKE_BUILD_TYPE=Debug >/dev/null 2>&1 && cmake --build targets/cjson/testbuild -j >/dev/null 2>&1 && ctest --test-dir targets/cjson/testbuild --output-on-failure"
+# Golden: full re-serialization + checksum of the corpus, byte-hashed.
+golden = "{binary} corpora/cjson/input/big.json 1"
+
+[corpus]
+manifest = "corpora/cjson/MANIFEST.sha256"
+root = "corpora/cjson/input"
+golden_sha256 = "corpora/cjson/GOLDEN.sha256"

targets/comrak/target.toml

@@ -18,8 +18,10 @@ license = "BSD-2-Clause"
 #   git clone <submodule repo> workspace/vendor/progit && git -C ... checkout <commit>
 
 [build]
-baseline = "cargo build --release --manifest-path targets/comrak/workspace/Cargo.toml"
-binary = "targets/comrak/workspace/target/release/comrak"
+baseline = "CARGO_TARGET_DIR={out} cargo build --release --manifest-path targets/comrak/workspace/Cargo.toml"
+binary = "{out}/release/comrak"
+sanitizer = "CARGO_TARGET_DIR={out} RUSTFLAGS='-Zsanitizer=address' cargo +nightly build --target x86_64-unknown-linux-gnu --manifest-path targets/comrak/workspace/Cargo.toml"
+sanitizer_binary = "{out}/x86_64-unknown-linux-gnu/debug/comrak"
 
 [bench]
 # Upstream workload (benches/bench.sh): full Pro Git book through the

targets/tokei/target.toml

@@ -9,8 +9,10 @@ commit = "fa44e5194060305576514d59b850353643afbfc8"   # v14.0.0
 license = "MIT OR Apache-2.0"
 
 [build]
-baseline = "cargo build --release --manifest-path targets/tokei/workspace/Cargo.toml"
-binary = "targets/tokei/workspace/target/release/tokei"
+baseline = "CARGO_TARGET_DIR={out} cargo build --release --manifest-path targets/tokei/workspace/Cargo.toml"
+binary = "{out}/release/tokei"
+sanitizer = "CARGO_TARGET_DIR={out} RUSTFLAGS='-Zsanitizer=address' cargo +nightly build --target x86_64-unknown-linux-gnu --manifest-path targets/tokei/workspace/Cargo.toml"
+sanitizer_binary = "{out}/x86_64-unknown-linux-gnu/debug/tokei"
 
 [bench]
 # Single rayon thread: measurement first (SPEC conventions); the