ВЕРСИИ (41)

- Merge mc-26.3 into main - Add a dev-only client gametest and build both jars in the publish workflow - Merge fix-lush-caves-prefilter into mc-26.3 - Skip biome columns only where a proof says the target cannot win - Check out shell scripts with LF line endings - Port to Minecraft 26.3 behind ClimateAccess and VersionCompat - Scaffold the 26.3 variant and decline 26.3 in the 26.1 jar

- Merge mc-26.3 into main - Add a dev-only client gametest and build both jars in the publish workflow - Merge fix-lush-caves-prefilter into mc-26.3 - Skip biome columns only where a proof says the target cannot win - Check out shell scripts with LF line endings - Port to Minecraft 26.3 behind ClimateAccess and VersionCompat - Scaffold the 26.3 variant and decline 26.3 in the 26.1 jar

- Add the deathMark config switch

- Show buttons and clickable coordinates only to viewers with the permission

- Give mark lines the same coordinate component as every other line

- Add /locate marks as the list command and click-to-teleport on mark lines

- Split mark and unmark into separate subcommands

- Suggest existing mark names and reserve the remove word

- Add the death mark and track/compass by mark name

- Add named marks and put every subcommand under /locate

- Make repeat poi searches near-instant and pipeline the disk reads

- Let the near ladder escalate to exhaustion or the time budget - Document the near radius as optional

- Cluster poi results into places and default the near radius

- Add async /locate poi with count and min_distance over all explored terrain

- Split hit presentation into its own class, add the pure-math test slice and run-both.sh

- Add biome pairs to /locatemore near

- Add min_distance to /locate counts and /locatemore near constellation search

- Add fabric-permissions-api support for /locatemore subcommands

- Tracker state dies with the server; battery gets its own port

- Results read as places: compass heading, pointer compass, teaching misses

- Ring-batch sorted frontier replaces the biome heap - Pipeline plumbing from the queueing analysis: wake, fence, side room - The depth window returns, proved this time

- The engine round: the exactness-critical trust decision, the ring-expansion pass and the whole presentation layer (boss bar, streaming, progress) now live in one place shared by the structure and biome engines, so improvements land everywhere at once - The biome prefilter gained a provable core: certified rejection margins computed from the parameter list and the noise functions' own bounds, valid on any datapack with no referee needed; the band between the intervals and the margins keeps the referee-guarded tier - The math memo survives its cap (two generations, promote on hit), chunk loads start the same tick they are requested, and biome columns are sampled in chunks - Releases now go through CI: one jar verified against both supported Minecraft versions on every push, published on tag

- Biome searches got two more layers of speed: a runtime-derived climate-interval prefilter rejects most columns on one cheap noise before the other four and every tree lookup (nearest desert: 688 ms to 75 ms on the release battery), and columns are sampled in chunks instead of one task each - Monument candidates fail on 8 corner samples instead of vanilla's full 3375-sample box on the miss path; the corners are members of vanilla's own sample set, so the rejection is provably permissive - Every result line gains a [track] button: live distance and a direction arrow in the action bar while you travel, self-clearing on arrival - nothing to clean up - The math memo survives its cap with a two-generation scheme, chunk loads start the same tick they are requested, and structure searches with modded placement types now fall back to vanilla instead of silently missing - All shortcuts remain referee-guarded: a depth-based shortcut that failed its referee during testing was withdrawn before release

- Biome searches run in parallel across the worker pool: batches of the nearest unsampled columns fan out while acceptance stays sequential over ascending distances, so exact order is preserved and full-radius sweeps drop from tens of seconds to a few - Five of the six climate functions are 2D in vanilla's router, so they are computed once per column instead of once per y sample. The shortcut is earned per world (startup probe at the world's height extremes) and watched by a standing 1-in-64 referee against the full sampler; any disagreement disables it for the session. Datapacks with 3D climate and non-multinoise dimensions take the plain path

- /locate biome gets the same two upgrades as structures: the search runs off the server thread (no tick freeze), and results come in exact distance order instead of vanilla's approximately-nearest spiral (the release battery caught vanilla reporting a desert at 4518 blocks with one at 4482) - Optional count: /locate biome minecraft:desert 5 streams the 5 nearest distinct patches with clickable teleports, y included - Hits within biomeSeparationBlocks (default 512) of an accepted hit are suppressed, so one giant biome cannot fill the list; search radius is biomeMaxDistanceBlocks (default 12800, double vanilla) - The exact_locate gamerule and improveVanillaLocate kill switch gate the vanilla one-result path exactly like the structure call sites

- Plain /locate structure (no count) now runs on the async engine too, exactly like the count variant: the command can never stall a tick. Before this, a worst-case query - a rare set variant kilometres away on a cold JVM - could freeze for seconds on the vanilla executor's synchronous path. /locate with and without a count are now the same thing, which is what everyone assumed anyway - With the locatemore:exact_locate gamerule off or the kill switch set, vanilla's executor runs untouched, as always - Eyes of ender and cartographer trades keep the bounded on-thread search (they are synchronous by contract and measure 9-900 ms); if a worst-case query ever exceeds its 3 s budget, vanilla's own search takes over - Single-result searches read naturally: "Searching for the nearest ..." and "Nearest ..."

- The synchronous budget behind the vanilla call sites (plain /locate, eyes of ender, explorer maps and trades) drops from 15 s to 3 s. Sized from measurement, not caution: the worst case forcible on a mature 536-region world - a monument trade map resolving under vanilla's skip-known filter with chunk generation on the thread - took 900 ms, so 3 s keeps 3x headroom over the worst observed while bounding a pathological stall at a fifth of the old ceiling. On the rare give-up the mod steps aside and vanilla's own search runs, and a session counter watches that this stays rare - The budget is now per-caller: the verify command and the release harness's generation-backed searches keep a generous budget, because they exist to run to completion

- The standing on-disk draw sample now reports in the server log (one INFO line per search that sampled: "On-disk draw sample: X/Y predictions matched already-generated chunks"). It counted before but full agreement was invisible behind the quiet chat summary, so silence was ambiguous between "all agreed" and "nothing sampled" - exactly the ambiguity a watchdog must not have. First field data: 10/10 matches against chunks generated by real play

An external architecture review of the whole mod ran after 1.8.0; this release ships everything it changed. - The region catalog no longer enumerates the region directory per search. That enumeration cost grew with world age and ran on the server thread for every vanilla call site (eyes of ender, cartographer trades, dolphins); it is now one memoized file-stat per region the search actually visits, with identical semantics. On mature worlds this removes a per-search cost that no fresh-world benchmark could see - The release battery now generates the multi-set answers for real (villages and fortresses through actual chunk generation) and diffs them against the trusted draw engine's answers within the same run, so a broken draw replica can never be blessed as a clean port again. A standing sample also compares draw predictions against chunks already generated on disk (drawSeen= in the summary) - the population no referee had ever watched - API v2, in its own package (com.rasmus.locatemore.api): SearchResult gains complete (false when a budget cut the search short - the difference between "only three exist" and "we ran out of time"), the chat-facing maxCount no longer silently truncates API calls, and config values are read-only to other mods - The draw fallback now survives datapack reloads correctly (keyed on the set's registry key instead of object identity), and a chunk whose structure data the parser no longer recognizes fails safely to a chunk load instead of silently falling back to math - The next mode is gone; SearchOptions.excludePrevious is the one way to skip previous hits. The summary line's avoided= figure now shows how many chunk loads the seed math replaced - The README's measured table was re-measured under the reproducible harness protocol (and A/B-verified against 1.8.0 first: the engine did not get slower, the old numbers were stale). Multi-set searches have collapsed since they were taken: villages 1.4 s to 0.17 s, fortresses 0.19 s to 0.04 s, cold, count 20

- Multi-member structure sets (villages, fortress/bastion) now resolve from seed math alone: the replica of generation's weighted draw between set members, which shipped in 1.7.1 as a measure-only referee, earned trust with 427 of 427 predictions confirmed by real chunk loads across 7 seeds. A fortress search that generated 15 probe chunks now generates none (307 ms -> 15 ms cold on the test seed; villages 1060 ms -> 44 ms), and the same trust covers plain /locate and every vanilla call site - Two guard rails ship with the trust: if any referee load ever observes a disagreement, that placement falls back to chunk loads for the rest of the session with a WARN, and sets larger than 8 members (datapacks) always resolve through real generation, because the replica was only proven against vanilla-scale sets - The release gate grew four lines: the golden battery's village search previously returned one hit plus four unresolved candidates on a freshly created world, because probe loads race the spawn regions being written. The draw verdict resolves them from math; all four recovered hits were ground-truthed by forcing their chunks to generate and re-reading the answer from disk

- Multi-member structure sets (like the set that shares placements between fortresses and bastions) now replicate world generation's weighted draw between members, so the engine predicts which structure a candidate chunk will actually produce. For now this is measurement only: every probe load doubles as a check of the prediction, reported as draw=hits/loads next to the math counter, and any disagreement is logged with the chunk and both structures. Search results are unchanged, and the release passed the full golden coordinate battery unchanged. Once the counter has proven the draw model in real worlds, a later release can use it to skip those loads too - fortress and bastion searches would then resolve from seed math alone - The mod description now leads with what the engine does for every search site instead of just the count argument

- Explorer maps now go through the exact engine: cartographer trades, treasure maps in loot chests, and dolphin treasure hunts. Candidates are walked in true distance order, pruned by seed math where provable, and exactly the first candidate vanilla's own skip-known filter and canBeReferenced accept is loaded - the reference mutation is vanilla's addReference, never reimplemented. The multi-second freeze these call sites are known for becomes a single chunk generation: measured 119-244 ms for mansion maps and 68-84 ms for buried treasure maps - The single-set math shortcut is deliberately disabled on the map path, because its filter must only ever be permissive: vanilla makes every accept/reject decision, the engine only orders the candidates and prunes the provably impossible ones - Functional proof in the release gate: running the map path twice from one origin returns the nearest, then the second-nearest (the reference was really taken); both rounds are part of the golden coordinate battery every release is diffed against

API: - Requests now queue (bounded, player searches always admitted first) instead of failing when every worker slot is busy - SearchOptions: per-call distance/time budgets clamped to the server's, allowChunkGeneration=false (SearchOptions.mathOnly()) answers what the math can prove and flags the rest instead of writing to the world, and excludePrevious makes 'find the next ones' a single call - LocateMoreApi.API_VERSION marks the contract Operators: - /gamerule locatemore:exact_locate (default true) toggles the vanilla call sites live and per world; the improveVanillaLocate config key remains the server-wide kill switch, and a startup log line makes the behavior change discoverable - Clean searches print a quiet one-number summary; the counters only appear when they explain chunk work, partial results or a referee disagreement - and a math-referee miss now logs a loud WARN with seed, dimension and chunk Leaner: - All development tooling left the jar: the sync/vanilla lab modes, apitest, the vanilla cache inspector (and its two mapping-fragile mixin accessors), the manual memo clear and the benchmark config key. Remaining operator commands: verify and prune - Every release is now diffed coordinate-for-coordinate against real generation on a fixed seed before shipping

- Removed the internal development bridge from the shipped jar entirely. It was disabled by default and dead on dedicated servers, but a released mod should carry no command-execution surface at all; development tooling now lives outside the mod

- Plain /locate, eyes of ender, and other mods calling vanilla's nearest-structure search now get the single-set math shortcut too: a cold mansion locate answers instantly, loads no chunks, and no longer leaves empty region files behind (vanilla's own scan path creates one per examined candidate in ungenerated terrain) - The sync and vanilla lab modes deliberately keep the old behavior as a generation-backed referee; at the same cold origin the referee returned the identical position while checking 303 candidates and creating 304 empty files, which is exactly what the trusted path avoids - /locatemore prune now also cleans the files vanilla's scans leave behind

- Removed the persistent negative index. An A/B measurement showed it saved zero time once the region catalog existed: the catalog already skips the disk scan for the ~95% of candidates whose region file does not exist, and scanning the rest costs nothing measurable. That deletes a save file, a fingerprint scheme and ~200 lines - The mod now keeps zero persistent state: its only cache is the in-memory session math memo, so nothing can go stale across restarts. /locatemore memo clear replaces /locatemore index clear - Orphaned structure_index files in existing worlds are ignored and harmless

- Searches for single-structure sets (mansions, ancient cities, monuments, temples, outposts, trial chambers and more) no longer load or generate any chunks: for a one-member set, generation would run exactly the math the search already ran, so the verdict is the answer. Cold mansion search: 11.7 s down to 0.16 s, with zero world-save growth - The math path now applies generation's full placement filter, including exclusion zones; outposts no longer waste probes on chunks generation refuses near village candidates - Multi-structure sets (villages, nether complexes) still verify through real generation, and a math=hits/loads referee in the summary line audits every such load. The shortcut shipped only after the referee measured 100% agreement across the single-set battery and async results matched the generation-backed sync mode position for position

- Fixed: chunk scans no longer create empty region files. Vanilla's scan path opens region files with CREATE, so every scan of a candidate in unexplored terrain wrote an empty .mca file, thousands per cold search. Each search now lists the region directory once and skips the scan for candidates whose region file is absent, which also removes the disk round trip for unexplored terrain - /locatemore prune deletes the zero-byte region files earlier versions left behind (run once per dimension) - The summary line now reports regionSkips

- Public API for other mods: LocateMoreApi.findNearest returns a future that completes on the server thread with the N nearest structures in exact distance order, on the same engine and budgets as the command - /locatemore apitest <structure> <count> runs the API end to end from in game - /locatemore index clear now also clears the session math memo, so admins have an invalidation path for datapack edits the fingerprint cannot see - The two internal budget checks no longer share a name

Vanilla /locate, eyes of ender, and any mod calling findNearestMapStructure now get the true nearest structure through the exact-order engine (fixes the MC-138887 class of wrong results; config key improveVanillaLocate, on by default, with automatic vanilla fallback on budget limits). Explorer maps keep the vanilla path. New: /locate structure <id> <count> next skips the structure you are standing in, and /locatemore verify <id> cross-checks the mod's disk parser against vanilla as a drift alarm.

Final pre-release hardening from the round-4 comparative audit: chunk futures acquired off the tick thread, snapshot height accessor for off-thread math, wider index fingerprint (biome/noise registries), honest partial-result reporting with unresolved-candidate counters, exception-proof async legs, benchmark modes gated behind config and admin permission, config clamps closing an overflow hole, and assorted smaller fixes.