Repository navigation
Description
The review of PR #4 confirmed that separate per-mode committed proof caches and per-root provisional results preserve initializer-cycle soundness. Stable failed proofs can still be revisited at repeated call sites.
Investigate caching failures with explicit causes (native, reflective, missing body, invokedynamic), without committing depth-exhaustion or recursion-dependent negatives that would introduce order-dependent precision loss. Add order/depth and initializer-cycle regression controls. Any performance acceptance must use representative requests that actually analyze multiple application graphs, reporting elapsed time and configured maximum heap plus sampled peak heap; single-graph timing is not acceptance evidence.
This is a nonblocking performance follow-up from #4 (comment), not an unresolved correctness claim or an asserted performance improvement.
Evaluation outcome
Evaluation completed; the stable-negative-cache candidate is not being adopted. This experiment did not establish a repeatable end-to-end performance benefit sufficient to justify the additional cache.
The local candidate records only explicit native, reflective, missing-body and dynamic-bootstrap failures, with separate proof modes and invocation keys. Depth exhaustion, recursion-dependent negatives and provisional initializer-cycle assumptions remain uncommitted. All 80 focused correctness tests passed, including order/depth and initializer-cycle controls.
Representative multi-graph experiment
- Real Netty change: BASE cf4657d to HEAD 7fc2dcb. Native Maven packaged inputs for common and buffer form one request with two separate graph scopes; both contain actual changed production classes.
- Baseline GraphFold:
2d842be48b99e71e150e56459ea93604cc467fba(the input fixes in Preserve signed archive evidence during static projection #9/Honor declared classpath order in static inputs #10). The candidate differs only in the dependency observation proof cache and its tests. Neither compiler outputs nor unrelated module inputs were merged into a fabricated graph. - Explicit selection:
PlatformDependent.hasUnsafe()Z = true. Each complete request applies BASE/HEAD folds with matched counts 26/26 for common and 49/49 for buffer (including 23 buffer-owned call sites per revision). - Six fresh JVMs in A/B/B/A/A/B order, three runs per variant, JDK 17 on macOS, two active processors, identical 8 GiB maximum heaps. The timed window is
engine.prepareplusengine.analyze; native archive preparation and report serialization are excluded. Heap usage is sampled every 100 ms plus endpoints. - All six complete reports are identical to the untimed qualified report, without diagnostic or node-ID normalization. All 42 input records and tracked source/class/resource/harness hashes remained unchanged.
| Metric | Baseline | Candidate |
|---|---|---|
| Median elapsed | 97.034 s | 95.268 s |
| Mean elapsed | 95.367 s | 95.250 s |
| Elapsed range | 91.634–97.432 s | 94.645–95.838 s |
| Configured maximum heap | 8 GiB | 8 GiB |
| Median sampled peak heap | 6.557 GiB | 6.541 GiB |
| Sampled peak heap range | 6.506–6.728 GiB | 6.515–6.693 GiB |
Median elapsed improves by about 1.82%, but mean elapsed changes by only 0.12%; one of three adjacent paired comparisons is slower, and the run ranges overlap. Sampled peak heap ranges also overlap. These observations do not establish a stable improvement or a lower required heap limit. No minimum-heap search or universal performance claim is made.
Scope of the conclusion
The request completes actual analysis and folding of both graphs, but its verdict remains FAIL: all four obligations are unproved, with 98,392 INCOMPLETE_GRAPH and two PRODUCTION_OUTPUT_CHANGED diagnostics retained. This is performance evidence for that completed, conservatively rejected request, not successful control/safety coverage or evidence about all applications. Cache-hit activity was not separately instrumented.
The investigation is complete with no production caching change. The local candidate and raw measurements are preserved outside the PRs; only this aggregate result is recorded here. Input-correctness fixes discovered during qualification remain independently reviewed in #9/#10.
Activity