Skip to content

v0.2.47

Choose a tag to compare

View all tags
@Armonte Armonte released this 16 May 01:48
· 74 commits to main since this release
v0.2.47
a2cb66b

v0.2.47 — Phase F partial fix: SOCD pre-apply on host before spec capture

What this fixes

Replay / spectator divergence caused by SOCD-mode-mismatch between the host and the spectator/replay. The host's GekkoNet-delivered raw inputs were stored verbatim into the spectator stream / .fm2krep file; the spectator re-applied SOCD using its own FM2K_SOCD_MODE env var. On any frame where the user held L+R or U+D simultaneously, the two modes resolve the conflict differently, the spec engine gets a different input than the host engine, and the divergence cascades.

The parity-diff dev tool had pointed at this: first divergence consistently in p1_pos / p1_script, NEVER in RNG. That pattern is "input was different, script chose a different action, position followed" — i.e. classic input-mismatch cascade, not a deterministic-sim leak.

Fix: pre-apply SOCD on the host side before sending to the spectator. The stored value now carries the host's resolution. Spec's subsequent SOCD pass becomes a no-op for the resolved bits, so the spec's local mode no longer matters. Facing-flip continues to run from each side's deterministic local sim state — that part was already correct.

What this DOESN'T fix

  • Older .fm2krep files (recorded pre-v0.2.47) captured RAW inputs. Opening one with a v0.2.47 spec that runs a different SOCD mode is still subject to the old mismatch. New recordings are safe; the older-replay desyncs that para reported may persist for their existing files.
  • "Spectator stuck on Connecting" — separate issue (NAT punch / port bind / snapshot path), not addressed here.
  • Replay desync past ~4000 frames may have additional sources beyond SOCD (facing-flip-by-game-state corner cases, uncaptured state, RNG drift on rare paths). If a fresh v0.2.47 recording still desyncs through parity_diff, the SOCD fix wasn't the whole story and we'll have more digging to do.

How to verify

If you can record a fresh v0.2.47 match and replay it (especially with a character / playstyle that does L+R or U+D often — taunts, command-grabs that need full circle inputs, etc.), the spec / replay should now stay in sync where it didn't before.

Otherwise in this build

Everything from v0.2.42 → v0.2.46 (vsync soft-cap, per-game-input routing, UTF-8 manifest resilience, match-result queue on WS drop, SOCD picker working offline, controller hot-plug, UI polish sweep, etc).

Assets 3
Loading