v0.2.48

v0.2.48 — Phase F: replay determinism fix (intermediate-frame RNG drift)

What this fixes

Every v0.2.43 desync report in the hub log had the same diagnostic signature: GameplayFingerprint forward != replay on every frame, RNG_Seed forward != replay on 3 of 5. The parity-diff tool corroborated: first divergence always in p1_pos / p1_script, never in RNG — a classic input/state cascade from a deterministic-sim leak.

The cause: rollback batches run [Load → AdvanceEvent → SaveEvent → AdvanceEvent → SaveEvent → ... → render-once-at-end]. Render fires at most once per batch, so PatchPostRenderRng only updates the LAST save's RNG with the post-render value. Every intermediate SaveEvent overwrites the previously-patched RNG with the post-sim value. Result: replay sim for frame K+2 starts from POST-sim-(K+1) instead of POST-render-(K+1). Render's RNG delta (R_{K+1}) is missing on replay. Per-frame, the divergence accumulates and eventually trips GekkoNet's desync detector.

This is the long-running "match desyncs mid-game, both players' games close immediately" bug that's been hitting pkmncc, WonderfulWorld, URORFG across multiple users since the v0.2.40 immediate-terminate behavior went in.

The fix: store POST-render RNG in a parallel buffer that SaveState_Save doesn't touch. Only PatchPostRenderRng writes to it, tagged with the frame number so stale values from a previous match don't replay. Before each AdvanceEvent's sim call, look up the stored POST-render RNG for the previous frame and overwrite the engine's live RNG with it. On the forward path this is a no-op (engine RNG already matches). On the rollback path it re-establishes the right starting RNG for each intermediate frame.

What this means for you

• You should stop seeing the "match ended unexpectedly, both games closed" mid-match desyncs.
• Spectator and .fm2krep replay should now stay in sync with the host's actual run for the entire match (combined with v0.2.47's SOCD pre-apply fix).
• This DOES NOT retroactively fix old .fm2krep files; those were recorded with the buggy RNG sequence.

How to verify

• Play a match. If you used to desync after a few thousand frames on a particular game (pkmncc, WW, URORFG, vanpri), it should not desync anymore — or if it does, the diagnostic in the upload pipeline will show a different signature and we'll know we're in new territory.
• Run stress mode locally (dev menu) — the GekkoStressSession rolls back every check_distance frames. Pre-fix this would desync within a few hundred frames; post-fix it should run for a long time without tripping.

Reminder for users still on v0.2.41

v0.2.41 has an unrelated FM2KHook.dll crash that was fixed in v0.2.42+. If your match closes WITHOUT a desync diagnostic in the log, you're hitting that, and need to click the "Update" pill in the launcher's top-right.

Notes carried from v0.2.42 → .47: vsync idle CPU/GPU soft-cap, per-game input override actually applies offline, launcher resilience against bad upload manifests, hook-side UTF-8 manifests, match-result queue across hub WS drops, SOCD picker actually applies, controller hot-plug, UI polish (refresh-id collision, Edit-button visibility, path slash normalization, doubled ellipsis, session-record on titlebar), SOCD-pre-apply on spec capture.