Apply ZacSharp's review of the Java pathfinder

[0Mattias]

ZacSharp reviewed cabaletta/baritone#5117 on Sep 14 and 15 and babbaj answered on Sep 16. This brings the fixes back from the fork's java-nether-pathfinder branch (0Mattias/baritone 1a3c2113 to 63a5813b, one commit per comment), the way #20 to #22 did for the earlier rounds.

• the x8 summary is exact on every clear and the exact flag is gone: no whole-chunk writer ever clears a block, so the scan the flag avoided never ran for them, and isEmptyX8 and isEmptyX16 can no longer answer "not empty" for an empty cube
• the "x8 is marked before its block" comment is gone; the class comment now says what keeps readers and writers apart, NetherPathfinderContext's lock, which every reader holds on the read side and every writer on the write side
• the summary is a byte[], one byte per section, as its comment already said
• insertChunkData and setChunkState are deleted; only tests called them, and the Oracle test helper now generates its chunks itself
• a chunk's origin is a final boolean fromCaller instead of a volatile int
• NodePos holds three ints instead of a BlockPos
• the set of chunks generated around is a fastutil LongOpenHashSet
• the package comment names the Minecraft classes the package uses

The pathfinder package and its tests stay byte for byte the branch's, modulo the package name and the licence header; NetherPathfinderContext and package-info.java take the same change by hand.

Also: this fork's own bugs, found by a survey of the codebase

Six Opus agents read the whole codebase before this round. Of what they flagged, these are the fork's own and are fixed here, each verified against upstream's 26.2 branch first:

• the pre-tick mixin's slice still named Yarn's MinecraftClient.itemUseCooldown; Mixin resolves an unmatched slice start as the start of the method, so the pre-tick event fired on the first gui read in tick(), before gui.tick() and the missTime write, instead of after it as upstream places it
• CalculationContext hardcoded the Frost Walker and Depth Strider levels to 0 since the Mojang-names migration; both now come from the enchantment components, as in upstream
• the cached-chunk obsidian rule compared a floor-relative y against -59 and could never fire; it fires for the bottom five layers of the Overworld and the Nether now
• every cabaletta/baritone link in the comments had become cabaletta/cheesecake, and one javadoc link pointed at a cheesecake.leijurv.com that does not exist; the two links a player can see point at this repo
• Fixer, a reflection scratch class from the migration, shipped in the jar; it is gone
• SETUP.md claimed a "remapped to intermediary" jar check that does not exist

Three findings are real but identical in upstream and are filed instead: #24 (a CALC_FAILED can be lost before the process sees it), #25 (special-block lookup y mismatch in the Overworld cache), #26 (MovementDiagonal tests the first corner twice for magma).

All 101 unit tests pass locally, and the in-world test passes in full on this branch with the fixes above in place: all twenty stages, the four elytra flights included, in two runs.

Second round: verifying this branch, and what it turned up

A sweep of the whole codebase (38 Opus finders and verifiers, then checked by hand) confirmed the commits above and found that f9d70ee, taken from upstream as it is, regressed pathing. Two in-world stages added here print the pathfinder's own costs for a one-block step:

step	before f9d70ee	after it	with the fixes
into water, no Depth Strider	9.091	4.633	9.091
onto a lava source, Frost Walker II	COST_INF	3.564	COST_INF
onto waterlogged stairs, Frost Walker II	3.564	throws	3.564
Frost Walker level with the boots on	0	2	2

Fixed, one commit each:

• the Depth Strider multiplier starts at 0: upstream's 1.0 is Depth Strider III and priced a step into water like a step on land for everyone without the enchantment (upstream 26.2 has the same value; recorded in the README)
• both canUseFrostWalker overloads are upstream's, a water source only and the execution side reading the enchantment; the fork's accepted lava and threw on waterlogged blocks, dormant while the level was hardcoded to 0
• four upstream fixes that had been missed: Litematica schema version 7 (e66fdea5, with a unit test on the gate), the try/catch around the elytra goal (e545294d), BetterBlockPos kept out of the block entity map (76d6a98d), #path without a goal (511a2ae8)
• the tool data component instead of Item.class and instanceof Item, which every item passes (cfb7970e)
• PlayerUpdateEvent PRE after the vanilla player tick, as upstream: at the head of the tick the bot's aim became the player's previous rotation, and the camera swept between the aim and the view on every frame under free look; the new stage saw the previous yaw a quarter turn from the camera's on 39 of 40 ticks
• an agentApiPort above 65535 no longer crashes the client on every launch, and a status request over the socket outside a world answers instead of throwing
• the cache's dimension rules key on the dimension id again, as upstream: the End's dimension type has skylight in 26.2, so the flag proxy read cached End blocks as stone and the bedrock rule from 59dcb00 fired in the End, contrary to that commit's message
• #find searches around the player's z, not its y, and spawner avoidance asks the cache for spawner, not mob_spawner; both inherited from upstream and recorded in the README

The README's verification section and the passEvents comment are brought up to date, and the inventory-pause stage can run on its own. 103 unit tests pass, and the two new stages, which fail on cdf2238, pass with the fixes.

🤖 Generated with Claude Code