perf(stack): reduce resource, chunk and actor stalls

[milesalexg-tech]

Resource-pack preparation could block connection threads, completed conversions accumulated without a retention limit, and chunk/actor work repeated while its inputs were unchanged. This change bounds background work and caches, skips irrelevant chunk cells, and reuses stable render data.

Changes

• Move source ZIP loading, decryption, persistence, and definition construction onto bounded preparation workers. Hash and decode completed network downloads there as well, count missing chunks in constant time, and release ZIP inflater resources immediately. Keep storage publication and resource advertisement on the owning event loop, with stale tracker/storage/channel guards and explicit failure handling.
• Separate in-flight conversion sharing from a completed LRU capped at 32 entries and 128 MiB of archive bytes plus estimated metadata. Bound converter/preparation queues, reject saturation without caller-thread fallback, and complete pending futures during shutdown. Preserve caching of unused downloaded INFO packs.
• Classify palette entries once, skip sections without connected states, and limit neighbor refreshes to candidate cells on the relevant border. Preserve waterlogging and upper-door behavior.
• Coalesce each connection's queued/executing chunk tick with a token-owned reservation. Release it on completion or rejection and retain dimension-replacement guards.
• Fix private-profile preparation on macOS: canonicalize ancestor aliases while rejecting owned-root symlinks, copy profiles with portable filesystem APIs, and inspect exact process environments through Darwin APIs. Retry verified process-lifetime races without exposing arguments or environments.
• Cache resolved actor draws across unchanged snapshots with a 1,024-entry LRU. Build a lazy effect-locator index once per accepted resource generation. Keep animated poses, UV expressions, scale, and yaw at their existing per-frame evaluation points.

The changes are folded into their owning upstreamable patches. Later patches are replayed with selected subpacks, builtin resource layers, custom conversion profiles, doors, and replaced-dimension checks preserved. Each owning patch includes updated review and testing context.

Validation

• Full ViaBedrock build and all Checkstyle tasks: 1,155 tests passed, 30 skipped, zero failures/errors.
• Full add-on build against the pinned ViaFabricPlus dependency and updated ViaBedrock library: 784 passed, 120 skipped, zero failures/errors; access-widener validation passed.
• ViaProxy build: 4 tests passed. CubeConverter dependency build: 18 tests passed.
• All 1,381 ViaBedrock classes/resources match both downstream packages byte for byte.
• Clean full-stack replays reproduce the exact source trees used for the successful builds.
• bun run pr check viabedrock applies the first upstreamable patch to the pinned upstream base without setup/deferred patches. That standalone checkout passes all 45 tests and both Checkstyle tasks.
• bun run check and git diff --check pass. Full Bun suite on macOS: 206 passed, 4 skipped, zero failures. Six Python process-inspection regressions pass. The platform changes resolve the 18 original macOS failures; repeated process-churn checks exercise transient Darwin stack-copy races.
• Focused regressions cover queue saturation/shutdown, cache eviction and oversized sharing, unused source downloads, stale publication, differential palette/border selection, tick reservation ownership, snapshot invalidation/LRU, and differential locator indexing. A 144-frame snapshot case resolves once; repeated locator requests across two geometries parse the model once.

No new live-server frame-time, FPS, or joining-time measurements were made. Completed-download inflation now runs off the event loop; client resource reloads and other workloads still need live profiling. These changes do not claim to eliminate every possible lag spike.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

---

_{Comment @coderabbitai help to get the list of available commands.}

[AlexProgrammerDE]

Manually integrated the reviewed resource, chunk and actor optimizations into main in eb94fe6. The commit and owning feature patches credit @milesalexg-tech as co-author.

Included:

• Bounded resource workers for preparation and completed download hashing/ZIP decoding, with current-session publication guards and shutdown completion.
• Constant-time download receipt accounting and immediate ZIP inflater cleanup.
• Palette classification reuse and candidate-only connected shape scans, including border and upper-door behavior.
• Resolved actor snapshot caching and an effect locator index with resource invalidation.

Preserved the current four-entry/64 MiB conversion cache limits, weak chunk reservations, weak actor ownership, sound sharing and cursor crafting fixes. Each change stays folded into its existing owning patch.

The macOS profile/process changes remain separate because they overlap existing workspace edits. This is a partial integration, so the PR remains open.

Validation: all four dependency builds pass, with 1,967 tests passed and 150 skipped. The standalone cache patch replays on pinned upstream and passes 45 tests plus both Checkstyle tasks. TypeScript and patch whitespace checks pass. No new live frame-time measurements were made.