perf(jit): reuse native timing storage and settle Tier 2 benchmarks

CHANGELOG.md

@@ -22,7 +22,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - JIT M3: cache one idle native entry handle per runtime with epoch-based
   invalidation (ABI minor 20), remove native-return queue allocation, and reuse
   CALL feedback storage. Reuse validated bytecode boundaries during helper PC
-  checks. Add acquisition metrics, lifecycle regressions, and
+  checks, and reuse one native execution timing stack across functions.
+  Add acquisition metrics, lifecycle regressions, and
   paired generic-call diagnostics. See `docs/M3.md`.
 
 ### Deprecated
@@ -31,6 +32,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 
+- Benchmark forced Tier 2 direct calls only after the required caller/callee
+  publications settle; reject compilation during the timed batch and refresh
+  stale local workload checksums.
+
 ### Security
 
 ## [0.12.2] - 2026-07-27

benchmarks/README.md

@@ -26,6 +26,16 @@ The default policy is five warmups, 30 fresh-process latency samples, and ten
 one-second throughput windows per workload and mode. The JSON records those
 actual values. Do not use reduced runs for performance claims.
 
+
+Forced Tier 2 probes can declare `globalThis.tier2ReadyInstalls` to require a
+known publication count before timing. The direct-call probe requires four:
+two baseline and two optimizing artifacts. An unrelated blacklisted function
+cannot bypass that requirement. Both optimized modes execute quiet warmup
+iterations while draining compilation; an explicit forced-Tier2 requirement
+also rejects unsettled startup or new compilation during the timed batch.
+Unsupported probes without the declaration retain bounded interpreter/Tier 1
+fallback. Use the same harness revision for both sides of a runtime comparison.
+
 For a fast functional smoke run of the focused scalar-loop, call-heavy, and
 property-heavy cases: