v0.12.0 — NeoForge 1.21.1 support + data-integrity hardening

BetterAutoSave now ships from a single source tree as both a Forge 1.20.1 and a NeoForge 1.21.1 build. The async-save core (scheduling, the relay state machine, workers, diagnostics) is zero-Minecraft-dependency pure Java, source-merged and shared by both loaders. This release also fixes the silent loss of chunk capability data (#8) and the data-integrity issues found in a pre-release code review.

NeoForge 1.21.1

• For NeoForge 21.1.x, Java 21+, server-side only; jar: shinoyuki_betterautosave-neoforge-0.12.0.jar
• Features, config, commands, and compatibility modes are identical to the Forge 1.20.1 build
• Serialization and disk-IO glue verified point-by-point against decompiled 1.21.1 sources; all 7 mixin injection points survive; entity region IO moved into SimpleRegionStorage in 1.21.1, adapted through a two-hop accessor so the async entity pipeline is reused as-is
• Async chunk save now includes NeoForge data attachments and the LevelChunk auxiliary light field, matching the vanilla on-disk format — mods using data attachments will not lose data during async saves
• Verified on a real NeoForge 1.21.1 server: all mixins load, periodic and operational /save-all run without main-thread stalls, workers drain cleanly on shutdown, restart loads with zero corruption

Data-integrity fixes

• Silent chunk-capability data loss (#8): in the default PARTIAL mode, the main-thread hand-built chunk NBT omitted the ForgeCaps block Forge injects at the end of ChunkSerializer.write, so any mod storing data in a chunk capability (e.g. illusion) lost it on every async save and lost it world-wide after a normal restart. PARTIAL/DISABLED now serialize ForgeCaps on the main thread, matching the vanilla format (the NeoForge data-attachment path was already correct)
• SavedData async-write race: a mod's save() may return a live internal tag; while a worker serializes it asynchronously, the mod mutating it can throw ConcurrentModificationException or write a torn NBT. The outer tag is now deep-copied before hand-off, matching NeoForge vanilla SavedData.save's own copy
• Data loss on degraded fallback: when workers all die from a systemic failure (typically OOM) and the pipeline degrades, queued tasks that had already declared their data "saved" but were never written used to be skipped by the shutdown flush and lost. The degrade transition now drains the residual tasks — chunks restore their unsaved flag for the vanilla synchronous fallback, SavedData is re-marked dirty, and entity losses (no coordinate recovery) are reported — all thread-safe

Other hardening

• The in-flight serializing counter is compensated under extreme queue OOM, so shutdown no longer burns its full timeout and falsely reports unsaved data
• SavedData atomic writes clean up the temp file if any step fails, leaving no orphans
• Operational /save-all flush's intentional deviations from a few vanilla side-effects are now documented; behavior unchanged

Upgrading

Drop-in jar replacement, no config changes. Forge users on 0.11.x are strongly encouraged to upgrade — 0.11.x still carries the #8 chunk-capability loss risk