Chunksmith

The first release since 3.4.1. Versions 3.5.0 through 3.12.0 were built and tested in-house but never published, so everything they contain arrives here -- this entry covers the whole line. The short version: a pre-gen could hold the entire world open until the server died, and that is fixed; throughput on a multi-core server is up substantially; and the mod now tells you what it is doing instead of leaving you to infer it from tick times.

Fixed

• A pre-gen could hold the entire world open, and the server would die to the tick watchdog. This is the headline fix. A held chunk does not cost one chunk: its ticket sits at FULL level and vanilla's distance manager propagates that level outward one ring at a time, so a single held ticket keeps a whole neighbourhood resident. Measured on a live 1.21.11 server, 20 held tickets held 3,507 resident chunks -- roughly 25 resident chunks per held ticket at pre-gen clustering. 3.4.1 had no cap on the settle frontier at all, which is how a real server reached 75,045 resident chunks, pinned its heap at 107% of an 8 GB -Xmx, and collapsed to 5 chunks per second with a 42-hour ETA. The frontier is now bounded by pregenSettleMaxHeld (default 256), and memory is bounded directly by throttleMaxHeapPercent rather than by any proxy.

• The unload backlog outlived the task that created it. Housekeeping was armed by a running task, so when a run paused or finished, whatever backlog remained was handed to vanilla's budgeted pass -- which does almost nothing once the tick is already over budget, and the tick was over budget precisely because of the retained chunks. The backlog was then permanent until a restart. Housekeeping is now armed while a BACKLOG exists, not while a task exists.

• Three separate mechanisms were suppressing their own recovery. Releases are driven by new chunks arriving, so gating dispatch also stopped the releases that would have opened the gate; a gate holding dispatch was given the small unload budget, so nothing unloaded and it never reopened; and the settle frontier froze at its cap during a hold while holding the very tickets that blocked recovery. All three now release, drain and re-arm from the tick instead.

• /cs debug answered "An unexpected error occurred trying to execute that command". The residency snapshot contained a literal percent sign and the message path runs through String.format.

• /cs continue reported "Task already started!" while the previous run was still draining, leaving the operator with a stopped pre-gen and a message saying the opposite. It now says what is actually happening.

• The ticket diagnostics were compiled into the shared mixin but written against 1.21.11-and-newer shapes, so this whole line of work could only build for 1.21.11. They are now gated on the presence of those shapes and every supported version has a jar again.

• A 26-line drift: ChunkPos became a record and lost its packed-long constructor.

• Config range-clamp warnings went nowhere. GsonConfig and TaskScheduler logged through java.util.logging, which no loader routes into the game log, so a clamped value was silently corrected with the explanation discarded. Both now use slf4j like the rest of the mod.

Added

• dispatchMaxConcurrent -- how many chunk requests stay in flight, and the single biggest throughput lever in the mod. It was previously fixed at 50 and reachable only through an undocumented chunksmith.maxWorkingCount system property. Nothing was saturated at that cap: CPU sat at 255 of 800 percent, workers at ~24 percent, the server thread at ~10. Measured on an 8-core dedicated server: 50 -> 31.6 chunks/sec, 200 -> 43.9, 600 -> 42.4. So 200 is the knee, worth +39 percent, and beyond it the remaining ceiling is vanilla promoting about 2.2 chunks per tick at 20 tps. The default now scales with the machine -- min(400, max(50, cores * 25)) -- with a floor of 50 so no server gets slower than it was. Live-settable with /cs set.

• A pre-gen now pauses when the server cannot sustain it, and resumes when it can. autoPauseOnOverload (default true) and autoPauseGraceSeconds (default 120). A gated run on an overloaded server does not stop, it stutters -- measured at 60 chunks in two minutes, which is indistinguishable from a hang, keeps the server under load throughout, and produces nothing. Both directions require the condition to hold continuously for the grace period. A human /cs pause outranks it in both directions and is never auto-resumed.

• throttleMaxHeapPercent (default 85) bounds the run by measuring memory. Three earlier releases tried to bound a pre-gen with a proxy -- queued writes, LOD queue depth, resident chunks, chunks added since the run started -- and each was wrong on a live server in a different way. A chunk is worth wildly different amounts of heap depending on what came with it, so no chunk count means the same thing on two worlds.

• The mod now reports what it is doing. /cs debug prints a chunk-residency snapshot, a ticket ledger, and an unload breakdown with a plain-English verdict; a drain logs one line when it starts and one when it ends, naming the outcome and how many chunks it freed, at WARNING level when chunks are left behind. Previously the only way to read residency at all was to set a low cap, start a run, and read the backpressure line -- which perturbs the thing being measured.

Changed

• throttleMaxAddedChunks now defaults to 0 (off). On a live server it closed at 22,000 chunks with the heap at 40 percent and stuttered a healthy run down to 60 chunks per two minutes. A pre-gen's resident set is its FULL chunks plus the mandatory worldgen context ring around each of them, and that frontier's perimeter legitimately grows with the radius. It remains available as an expert knob.

• The tick-budget baseline now requires 15 consecutive idle ticks, and probes every 60 seconds instead of 120. A one- or two-tick gap between chunks is not idle -- it is still paying for the chunk that just landed -- and sampling those taught the baseline the mod's own aftermath, observed as a baseline reading 49ms and then 116.8ms with no change in load.

• The LOD startup message no longer claims LOD generation costs ~16 percent of pre-gen speed. It does not: measured over matched windows, 36.1 chunks/sec with LOD on against 34.2 with it off. The message now reports the measured cost as none.

• The README documents nine config keys that previously had no documentation anywhere.

Removed

• The stale-ticket purge, and its purgeStaleVanillaTickets and staleTicketThreshold keys. A controlled A/B put residency at 16,800 with it on and 16,820 with it off -- a 0.1 percent difference -- while it evicted 10,765 tickets. It recovered nothing and was removed rather than kept as reassurance.