Bootstrap an isolated daemon for real-model memory evals and always clean up

evals/memory/BUILD.bazel

@@ -7,6 +7,7 @@ py_test(
         "//benchmark_data:run_benchmarks.py",
         "//evals/hermetic:score.py",
         "//scripts:build-eval-binary.sh",
+        "//scripts:lib/portable_process.py",
     ],
     imports = ["."],
     legacy_create_init = 0,

evals/memory/README.md

@@ -30,6 +30,12 @@ python3 evals/memory/run_real_model.py --scenario memory-no-pollution \
 The real-model runner creates a unique temporary user home for each run and
 installs the normal user-level Hermes integration there. It never selects a
 named Hermes profile or uses `HERMES_HOME` as TraceDecay storage.
+Each scenario also owns an isolated `tracedecay daemon run` (private socket,
+started via `scripts/lib/portable_process.py`) that must be ready before
+`init`. Setup, daemon, agent, and cancellation failures are written to the
+report as error entries. Cleanup is attempted on every exit path. If owned
+work cannot be stopped or deletion fails, retained paths and cleanup errors
+are recorded in the report.
 
 Run reports land under `evals/memory/runs/<timestamp>/` (gitignored).