feat: add secure-runtime snapshot comparator

benchmarks/real_world/README.md

@@ -82,6 +82,51 @@ HTTP/WebSocket claims addressable by this tool. The output is not a score, and a
 single-review or exploratory set must not be presented as canonical benchmark
 truth.
 
+## Secure AST/runtime comparison protocol
+
+`compare_runtime.py` compares positive observations from isolated runtime
+introspection with secure static results. Runtime output is a **comparator, not
+truth**: `secure_only` and `runtime_only` are disagreements for source review,
+not false positives or false negatives.
+
+A target/baseline comparison consumes four independently generated schema-v1
+artifacts. All four must declare the same app/bootstrap/app-variable/backend
+selection. Runtime records currently require null `app_entry` and
+`bootstrap_entry`, because isolated runtime does not yet execute those options.
+Target and baseline may use different dependency environments, but the secure
+and runtime records within each snapshot must attest the same source, tool,
+lock, immutable runtime image, and SBOM. Runtime provenance additionally binds
+the seccomp policy, sandbox policy, and exact effective invocation. Provenance
+is strictly mode-specific and its canonical digest is retained in the report.
+
+```bash
+python benchmarks/real_world/compare_runtime.py \
+  --secure-target /results/secure-target.json \
+  --runtime-target /results/runtime-target.json \
+  --secure-baseline /results/secure-baseline.json \
+  --runtime-baseline /results/runtime-baseline.json \
+  --output /results/secure-runtime-comparison.json
+```
+
+The report keeps four-artifact operational evidence (success/abstention,
+structured failure phases, inventory strength, timing, and peak RSS) separate
+from the paired-success quality subset. Timing has exact `list` and `impact`
+measurement states; resources have an exact `peak_rss_bytes` state. Each state
+is either `measured` with its unit or `not_measured` with a reason, matching the
+benchmark telemetry convention. Secure inventory strength is exactly one of
+`established`, `conditional`, or `unavailable`; runtime strength is exactly one
+of `runtime_observed`, `runtime_conditional`, or `runtime_unavailable`. A
+successful pair contributes quality only when both timing states and peak RSS
+are measured for both modes. Inventory and impact disagreements are reported
+exactly and with normalized path parameters;
+target/baseline lifecycle deltas are computed separately for secure and runtime
+modes. A failed artifact never contributes partial quality metrics. This tool
+compares one matrix; it does not aggregate or publish a corpus. Comparison
+output is no-clobber and cannot target source artifacts, frozen corpus/review
+files, canonical ground truth, or `benchmarks/results/**`, including through a
+symlinked parent. Its secure publisher currently requires the same Linux/POSIX
+filesystem semantics documented above.
+
 ## Frozen 50-project expansion
 
 Issue #103 adds a disjoint, metadata-only expansion with 50 projects and 100