This commit is contained in:
bryanthaboi
2026-07-28 15:01:59 -04:00
parent 04ece522d7
commit a03f69926e
11 changed files with 1081 additions and 17 deletions
+63 -4
View File
@@ -112,6 +112,19 @@ local function scalar(v)
return tostring(v)
end
-- A volatile slot is "off" three interchangeable ways -- absent, false, or
-- a counter sitting at 0 -- because every reader tests `if b.key then` or
-- `key > 0`. They have to hash the same, or a match ends over a difference
-- that does not exist. They routinely disagree: the menu-phase flinch
-- clear (BattleState:update) and the FIGHT-branch boundTurns mirror
-- (fightLockedAction) are written by whichever machine is sitting at ITS
-- OWN menu, and in a link battle that is a different battler on each side,
-- so one peer holds flinched=false / boundTurns=0 where the other still
-- holds nil with two identical simulations underneath.
local function off(v)
return v == nil or v == false or v == 0
end
local function stageStr(b)
local out = {}
for i, stat in ipairs(STAGES) do
@@ -131,7 +144,7 @@ end
local function volStr(b)
local out = {}
for _, key in ipairs(VOLATILE) do
if b[key] ~= nil then
if not off(b[key]) then
out[#out + 1] = key .. "=" .. scalar(b[key])
end
end
@@ -168,6 +181,17 @@ end
local PARTS = { "actives", "volatile", "bench" }
-- Which components are allowed to end a match. `actives` and `bench` carry
-- what decides one -- species, HP, status, stat stages, PP, the rest of the
-- party -- so a divergence there is a real split between the two
-- simulations and stays a draw. `volatile` is per-turn bookkeeping that
-- both sides recompute from the authoritative state every turn (see `off`
-- above): it can disagree for a turn without either side being wrong, and
-- when it does mean something real it lands in `actives` as damage or
-- status within a turn or two, where it is caught. Ending a match on it
-- alone cost players games they were winning, over nothing.
local FATAL_PART = { actives = true, bench = true }
-- opts: { myParty = packed, theirParty = packed, theirName, role =
-- "host"/"guest", seed, verdict, strict }. Returns nil plus a reason when
-- the handshake says the two link surfaces don't match: a lockstep
@@ -315,12 +339,23 @@ function LinkBattle.new(game, net, opts)
Logger.warn("link: desync turn %s component=%s (%s vs %s)",
tostring(turn), component, tostring(localH), tostring(remoteH))
Runtime.emit("link.desync", { turn = turn, component = component,
localHash = localH, remoteHash = remoteH })
localHash = localH, remoteHash = remoteH,
fatal = true })
endAsDraw(s, Strings(
"Link desync!\n%s differs.\fAre both games\nrunning the same\nmods?",
component))
end
-- a non-fatal component split: both sides log it and carry on, so the
-- match is decided by the battle rather than by bookkeeping
local function noteDrift(s, turn, component, localH, remoteH)
Logger.warn("link: %s drift on turn %s (%s vs %s) -- match continues",
component, tostring(turn), tostring(localH), tostring(remoteH))
Runtime.emit("link.desync", { turn = turn, component = component,
localHash = localH, remoteHash = remoteH,
fatal = false })
end
-- a verified turn stays recorded: consuming it here left a finished
-- battle holding 0-1 entries, so the whole-battle sweep the link suite
-- runs over localHashes had nothing left to compare
@@ -333,8 +368,11 @@ function LinkBattle.new(game, net, opts)
if mine and theirs then
for _, component in ipairs(PARTS) do
if mine[component] ~= theirs[component] then
reportDesync(s, turn, component, mine[component], theirs[component])
return
if FATAL_PART[component] then
reportDesync(s, turn, component, mine[component], theirs[component])
return
end
noteDrift(s, turn, component, mine[component], theirs[component])
end
end
end
@@ -565,10 +603,19 @@ function LinkBattle.new(game, net, opts)
if net.closed and not s.linkEnded and not s.result then
endAsDraw(s)
end
-- Both early returns below skip baseUpdate, which is where the
-- presentational clock normally advances -- so tick it here, in the same
-- order baseUpdate would (fx, then the queue). Without this an
-- animation caught mid-flight froze for the whole wait: a flash stopped
-- on its inverted BGP step and repainted the UI in inverted shades, and
-- a pic part-way through a slide or grow-in stayed off screen, which is
-- the "the screen went inverted" / "a Pokemon just vanished" pair.
if s.phase == "waitRemote" then
s:tickFx()
return -- the other side is still choosing
end
if s.phase == "messages" and s.afterQueue == "linkNext" then
s:tickFx()
if not s:updateQueue() then
s.afterQueue = "menu"
s.phase = "menu"
@@ -742,6 +789,13 @@ function LinkBattle.newSpectator(game, net, opts)
end
s:act(function()
-- the two real players cleared their flinch flags when their move
-- menu opened; a spectator has no menu, so it does it here instead,
-- at the same point in the turn (see BattleState:clearTurnFlinches).
-- Without this a flinch survived into the next turn and ate a move
-- that landed in the real match, and the replay -- sharing the RNG
-- stream -- was watching a different battle from that point on.
s:clearTurnFlinches()
local hostAction = hostMsg.kind ~= "switch" and hostMsg.kind ~= "run"
and decodeWireAction(s, hostMsg, s.player) or nil
local guestAction = guestMsg.kind ~= "switch" and guestMsg.kind ~= "run"
@@ -837,7 +891,11 @@ function LinkBattle.newSpectator(game, net, opts)
if net.closed and not s.linkEnded and not s.result then
endSpectate(s)
end
-- same as the real-participant loop: every path that skips baseUpdate
-- still has to advance the presentational clock, and a spectator sits in
-- waitBoth between every single turn
if s.phase == "messages" and s.afterQueue == "linkNext" then
s:tickFx()
if not s:updateQueue() then
s.afterQueue = "waitBoth"
s.phase = "waitBoth"
@@ -845,6 +903,7 @@ function LinkBattle.newSpectator(game, net, opts)
return
end
if s.phase == "menu" or s.phase == "waitBoth" then
s:tickFx()
return -- frozen between resolved turns; never a real decision here
end
baseUpdate(s, dt)