PathWeaver

0.5.1 — The half-fixed race

Hotfix for 0.5.0, published the same day. Two independent reviews of the shipped code found a blocker and a behavioural bug that eight earlier rounds had missed.

Fixed

• Mob.maxUpStep() was still being called from worker threads by four mob families. 0.5.0's WalkNodeEvaluatorMixin redirects the two calls inside WalkNodeEvaluator and its javadoc declares the attribute race eliminated. It was eliminated for two of the six admitted families. AmphibiousNodeEvaluator overrides getNeighbors and makes its own maxUpStep() call at offset 72 of its own bytecode, and a mixin transforms only its target class — so axolotls, turtles, drowned and frogs kept racing a read-modify-write on non-volatile attribute state from inside the A* loop. require = 1 passed on WalkNodeEvaluator and said nothing about the subclass.

  Fixed by AmphibiousNodeEvaluatorMixin, and guarded generally: LiveMobReadCoverageContractTest now scans every allowlisted evaluator for live-mob reads it declares itself and fails unless a mixin exists for that exact class. A seventh evaluator, or Mojang moving a call into a subclass, now breaks the build instead of going quiet.

• A recompute discarded a movement request the caller had been told succeeded. The supersede in pathweaver$supersedeBeforeRecomputeGuard rolls the optimistic targetPos back to its pre-dispatch value — and vanilla reads targetPos two bytecodes later to decide where to recompute. So a mob that had just accepted a move to A re-pathed to the destination it had abandoned, silently. The claimed destination is now preserved across the supersede, pinned by the routing game test.

• The startup report printed a false family count. "PathWeaver is doing NOTHING on this pack. All N movement families…" printed the size of the denial set, not the number of families — and isDenied matches with isAssignableFrom, so denying WalkNodeEvaluator alone blocks five families while the set size says one. A pack denying only Swim announced that nothing was running while five families dispatched. It now asks the gate per family and reports partial states.

• The land-registry diagnostic counter is a LongAdder rather than an AtomicLong. Fabric injects that lookup once per block position a search examines, so every worker in the pool was contending on a single cache line thousands of times per search — costing worker throughput, which surfaces as a worse install ratio.

Documentation

• DESIGN.md §9 carried a 0.2.2-era micro-benchmark reporting the mod as a net loss (2.927 ms OFF vs 3.012 ms ON). It described the rejected engine of §8, was labelled "not a user-real benchmark" with "noisy pairs", and sat unchanged for three releases — long enough that a senior reviewer read it as the project's performance evidence and concluded the mod does not pay off. Replaced with measurements from the shipped jar, including what they do not establish. §1–§7 are now marked as a 0.2.x record; they still claimed flying and amphibious were ineligible.
• EvaluatorCloner's no-arg guard was commented as though it refused to rebuild a subclass carrying its own state. It does not — the fallback reaches the same constructor — and it could not, since every land evaluator declares scratch fields. The comment now describes what it actually does and names the real limitation.
• New §13 records a known gap rather than patching it badly. A recompute-originated dispatch leaves the mob pathless for a tick, because vanilla nulls path before calling createPath. Returning the pre-null path was implemented, and broke the exact-Swim witness: a non-null this.path re-enables the vanilla reuse short-circuit and Feature B elision on a seam where neither may fire. The correct fix needs the request to carry its origin, which belongs in a release that can be soaked.

Measured

repathToleranceBlocks swept at 0/1/4/8, two runs each, order-reversed: no measurable effect on tick time. A first pass appeared to show 1 as clearly worst, which turned out to be a single outlier run inflating a mean over n=2.

268 unit tests, four server harnesses, the client harness, and an amphibious session run at repathToleranceBlocks=1: 28,648 dispatched, 28,625 installed, 0 cost drift over 288 checks.