Loot & Baloot

v3.2.10 (2026-05-15)

Full Changelog Previous Releases

• docs(CHANGELOG): add v3.2.10 release notes
• test(BQ): polish play-observer guard fixture/comments
  Comment/fixture-only polish on the v3.2.10 play-observer guard
  (no runtime logic change; harness stays 1320/0):
  1. Net.lua safeOnPlayObserved docblock: "chat + freezelog" was
     inaccurate — log("Error", ...) writes the B.Log ring buffer
     (L.Error("net", ...)), it does not directly print to chat.
     Reworded to "B.Log ring buffer via log("Error", ...) + the
     freezelog".
  2. BQ.1 host fixture: seat 2 (Mants) now holds {"7C"} in
     S.s.hostHands instead of an empty hand, so the play is a
     genuine legal follow-suit. The regression no longer leans on
     illegal-play tolerance — it isolates purely the
     observer-throw half-applied-stall it is meant to guard.
     Verification: full harness 1320/0 (unchanged), H1=11/0,
     H7=9/0, git diff --check clean.
     Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
• fix(Net.lua): guard play observer so human plays always advance
  Confirmed root cause of the user-reported recurring ~60s
  multiplayer freeze (Game 9M6KAP, host Mohtaal + Mants + 2 bots):
  The five play sites (_OnPlay, LocalPlay, _HostTurnTimeout AFK,
  MaybeRunBot success, MaybeRunBot recover) all applied the card
  via S.ApplyPlay — card becomes visible on the table, hand count
  drops — and THEN called the raw B.Bot.OnPlayObserved(...) BEFORE
  N.CancelTurnTimer() / N._HostStepPlay(). When the bot-memory
  observer threw, the play was visibly half-applied but the turn
  never advanced and the previously-armed AFK timer stayed live.
  The game "recovered" only when K.TURN_TIMEOUT_SEC (60s,
  Constants.lua:452) fired N._HostTurnTimeout and force-stepped
  the chain — exactly the reported symptom: "host shows 'Mants to
  act', Mants's card is on the table, resumes after ~60s". Not
  normal AFK; not a mere visual-refresh bug.
  Fix: new file-local safeOnPlayObserved(seat, card, leadBefore,
  context) that nil-checks B.Bot.OnPlayObserved, pcall-isolates
  the call, logs the error to chat + freezelog, and never
  rethrows. Bot-memory observation (void inference / firstDiscard
  / Fzloky / AKA dedup / trump-tempo) is a non-authoritative
  side-channel; its failure must never block the authoritative
  pipeline (send / cancel-timer / host-step / bot-dispatch / AFK
  recovery / UI refresh). All five raw call sites now route
  through the helper. Bot.lua observer logic is unchanged
  (investigation did not implicate a separate observer bug; the
  defect was the unguarded call ordering in Net.lua).
  Tests: new BQ section (19 checks) inside the AZ Net harness:
  BQ.1 (4) — host _OnPlay half-applied stall regression.
  Pre-fix BQ.1b FAILS ("turn ... got 2"), BQ.1d
  FAILS (rethrow); post-fix turn==3, no rethrow.
  BQ.2 (3) — non-host LocalPlay send-safety: SendPlay still
  broadcasts MSG_PLAY despite a throwing observer
  (real broadcast spy).
  BQ.3 (9) — source pins: all five sites route through
  safeOnPlayObserved + negative pins that no raw
  B.Bot.OnPlayObserved(seat/localSeat,...) direct
  call remains (sole raw ref is the helper's pcall).
  BQ.4 (3) — helper never rethrows; still calls through on the
  happy path.
  Verified pre-fix: BQ.1b/1d/2a/2c/3*/4* FAIL with Net.lua
  stashed; post-fix 1320/0 (was 1301/0). H1/H7 smokes clean
  (11/0 + 9/0). git diff --check clean.
  Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
• docs: clarify v3.2.9 playtest checklist marker refs
• docs: add v3.2.9 manual multiplayer playtest checklist