Optimize visual runtime and add upstream A/B validation

[TypeThe0ry]

Summary - optimize the production rewrite fixed at c28db0146ec2f35eaf066b4202b33f08d06bbc4b - reduce dropped-item visibility queue churn, bound text parsing caches, avoid empty drains, and clear retained manager/cache state on shutdown - remove avoidable block-state snapshots in furnace/beehive paths and tighten tile-entity lifecycle/index cleanup - preserve preference save/load ordering across quit/rejoin and plugin shutdown - add reproducible official-upstream vs rewrite runtime comparison with one prepare/download/build job and locally distributed, hash-verified artifacts - run independent restart-isolated samples in parallel; no fake cross-runner A/B pairing ## Correctness review No merge-blocking correctness issue remains in the fixed production candidate. Runtime optimization flags retain conservative production defaults; the benchmark requests them only in its controlled profile. ## GitHub validation All checks on fbcd0d569a65e3e46b161cd62151e79091e894e1 are green: - Java 25 build/test - A-B Performance Benchmark - Phase 2 Packet Capture A-B - Phase 2 Runtime A-B - Upstream Runtime Comparison smoke (12/12 jobs) - formal upstream runtime comparison (24/24 samples plus both aggregates and global gate) Formal run: https://github.com/EllanServer/InteractionVisualizer/actions/runs/29525891826 ## Formal stable-regime results Paper 26.1.2-74, 4 GiB symmetric heap, 20 TPS target, one protocol observer, 60 s warmup + 20 s settle + 120 s measurement. Each scenario has 12 independent restart-isolated samples (official A = 6, rewrite B = 6). | Scenario | Metric | Official median | Rewrite median | Median B/A | Bootstrap 95% CI | |---|---|---:|---:|---:|---:| | 512 dropped items | mean MSPT | 3.397374 ms | 2.460815 ms | 0.724329 | [0.545511, 0.918597] | | 512 dropped items | P95 MSPT | 30.005560 ms | 14.885620 ms | 0.496095 | [0.236532, 1.035402] | | 512 dropped items | P99 MSPT | 52.528233 ms | 17.202398 ms | 0.327489 | [0.242151, 0.440776] | | 256 active blocks | mean MSPT | 3.818482 ms | 0.941192 ms | 0.246483 | [0.212393, 0.283162] | | 256 active blocks | P95 MSPT | 4.495410 ms | 1.317267 ms | 0.293025 | [0.266592, 0.329043] | | 256 active blocks | P99 MSPT | 5.240943 ms | 1.567224 ms | 0.299035 | [0.262858, 0.354392] | Both scenarios pass the registered improvement and non-regression gates. A and B remain near 20 TPS, every formal sample keeps the observer online, and all 24 server processes exit cleanly. Global verdict: rewrite-better-across-both-scenarios. ## Stress evidence At 4 GiB with 1024 active blocks, official upstream A lost the observer to keepalive/packet-flood pressure in 6/6 runs; rewrite B completed 6/6. The failed upstream MSPT is not mixed into the stable CPU statistics because observer loss would underestimate its work. Stress run: https://github.com/EllanServer/InteractionVisualizer/actions/runs/29524047760 An earlier 2 GiB feasibility run also produced official-upstream OutOfMemoryError: Java heap space, while rewrite samples completed. Treat this as stability/memory evidence, not a retained-heap curve. ## Benchmark policy - build and download dependencies exactly once per campaign - distribute one internal tar artifact to measurement runners and verify the tar plus critical-file hashes - preserve the normal 20 TPS target - isolate server and protocol client CPU sets - use independent-sample bootstrap for parallel runners - require artifact/config/JVM/CPU/scenario/runtime-flag/TPS/observer/shutdown provenance before aggregation ## Caveat This is a synthetic single-observer GitHub-runner comparison, not a long multi-player production soak. A 30–60 minute multi-player retained-heap/JFR run remains a useful post-merge follow-up, but it is not a blocker for this PR.