Fix Block#getType returning STONE, ordered synchronous setBlockData, block-state properties, and bridge hot-path costs

[TR1LON]

Summary

Correctness and hot-path fixes on the block read/write path and event registration — complementing the j4rs→JNI and dedup work that just landed in 43e65ca (this branch is rebased on top of it; three further issues I'd found — the per-byte event marshalling, the response double-unwrap, and the block_in_place around the atomic health read — were already fixed there, so they're not in this PR).

1. Block#getType() could never return the actual block

PatchBukkitRegionAccessor#getBlockData called the bridge, then discarded the response and returned STONE for every non-air block:

var response = NativeBridgeFfi.getBlockData(request);
if (response != null && !response.getBlockState().isEmpty()) {
    return Bukkit.createBlockData(Material.STONE);   // response never read
}

Now feeds response.getBlockState() through Bukkit.createBlockData(String), which already handles namespaced keys, property suffixes and legacy names.

2. Block-state properties survive the bridge

get_block_data returned only block.name — stairs lost facing, doors lost half. Now includes properties via Block::properties(state_id).to_props() → minecraft:oak_stairs[facing=north,half=top]. Symmetrically, set_block_data parsed the string by splitting on [ and placing default_state; it now applies the [k=v,...] suffix through Block::from_properties(...).to_state_id(...).

3. Block#setType returned before the block was set

set_block_data was fire-and-forget (runtime.spawn): read-after-write raced, and two writes to one position could land in either order. Bukkit's contract is synchronous, ordered mutation. Now uses the same block_in_place + block_on pattern other callbacks already use, and honors applyPhysics (NOTIFY_ALL vs NOTIFY_LISTENERS). A comment documents the re-entrancy constraint for anyone wiring block-physics events to the JVM later.

4. ~10 µs of guaranteed ClassNotFoundExceptions per block read

PatchBukkitBlockData.newData ran Class.forName("net.minecraft.SharedConstants") and Class.forName("org.bukkit.craftbukkit.block.data.CraftBlockData") unconditionally on every call. Neither class can exist in this process (no NMS/OBC on the classpath), so both were guaranteed exceptions. Measured on JDK 21/25: ~10 µs per getType() call — ~95% of its total cost (the entire FFI round trip is ~450 ns). Removed.

5. Rust-side registration dedup (defense in depth)

Complements 43e65ca's Java-side registeredBridgeEvents: a (plugin, event type) guard in the Rust registration callback itself, so a duplicate Pumpkin handler can't be created even by callers that bypass PatchBukkitEventManager (direct HandlerList registration, plugin reloads). Each duplicate handler costs a full serialize + cross-thread round trip per event fire and re-invokes listeners that already ran.

6. Misc

PatchBukkitBlock allocated a metadata HashMap per instance — getBlockAt() creates a fresh wrapper per call, so every block read produced ~80 B of garbage. The map is now lazy.

Verification

Developed in a sandbox without Maven Central or a nightly toolchain, so verification is compile-level and unit-level rather than a live server run — CI should be the final arbiter:

• Java: the touched files + protoc 35.1-generated proto sources compile cleanly under JDK 25 (Temurin 25.0.2) against the classpath extracted from the nightly release binary; unmodified master compiles identically under the same harness (A/B baseline). Re-verified after rebasing onto a6ca90a.
• Rust: parse-level rustc check and rustfmt --check pass on the touched files. Every new API verified against current Pumpkin sources: Block::properties/to_props/from_properties/to_state_id (pumpkin-data generated block.rs), BlockFlags::NOTIFY_LISTENERS (pumpkin-world/src/world.rs). A full offline cargo check wasn't possible.
• Behavioural: the block-state string build/parse logic passes 6 standalone cases (property round trip, empty suffix [], whitespace, missing prefix).

Judgment calls worth a maintainer's eye: the sync set_block_state is a semantics change (fire-and-forget → blocking, matching Bukkit's contract); and getBlockData on an unloaded chunk now reports what Pumpkin returns for it rather than always STONE.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Nk1NUKr7NNGxaU3jW5Nmyz

[Snowiiii]

Ig LGTM, Thanks