Loot & Baloot

v3.2.18 (2026-05-17)

Full Changelog Previous Releases

• docs(CHANGELOG): add v3.2.18 release notes
  ISMCTS runtime memory guard — targets a whole-WoW-client freeze/crash
  class (high memory, ~trick 3, 1-human + 3-bot Saudi-Master games):
  memory emergency brake + per-move telemetry via /baloot ismctsdiag +
  conservative 3-bot world-count caps + single post-move incremental GC
  step (never a full collect). Also folds in the non-addon CLAUDE.md
  CurseForge project-ID doc fix. Single release point after v3.2.17.
  Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
• fix(BotMaster): v3.2.18 Codex polish — GC-step 0-disable + worldscompleted reset
  Two small follow-up blockers from Codex review of abfe653 (no rewrite
  of abfe653; follow-up commit only):
  1. GC-step disable semantics. Constants doc says
     K.BOT_ISMCTS_GC_STEP_KB = 0 disables the post-move step, but the
     bare if gcAvail then collectgarbage("step", (... or 256)) still
     issued step(0) when the tunable was 0. Now resolves
     local gcStepKB = (K and K.BOT_ISMCTS_GC_STEP_KB) or 256 and
     gates if gcAvail and gcStepKB > 0 then so 0 truly disables.
  2. Telemetry reset consistency. The per-move reset comment promises
     /baloot ismctsdiag never shows a stale world count on a
     short-circuit, but _lastWorldsCompleted was not reset there — the
     no-legal-moves / legal-build-failed exits could surface a prior
     heavy move's count. Added BM._lastWorldsCompleted = 0 to the
     reset block before the short-circuit exits.
     Tests (+8): G.3j/j2/j3/j4 pin the resolve + > 0 gate + gated call
     and assert the old un-gated form is gone; G.3k still enforces no
     collectgarbage("collect"); G.3m positionally pins the
     _lastWorldsCompleted reset before short-circuit exits; G.7 is a
     deterministic behavioral test that stubs R.IsLegalPlay→false to take
     the no-legal exit (which has no own _lastWorldsCompleted write) and
     proves a seeded stale 999 is cleared to 0 by the reset block.
     Full harness 1545/0; H1 11/0; H7 9/0; git diff --check main..HEAD
     clean.
     Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
• fix(BotMaster): ISMCTS runtime memory guard + 3-bot world-count cap (v3.2.18)
  HIGH stability incident: WoW client freeze / high-memory crash around
  trick ~3 in 1-human + 3-bot Saudi-Master games. The v0.11.17
  wall-clock budget bounds ISMCTS latency but not allocation — on a
  fast client the determinization/rollout loop allocates megabytes of
  transient tables well within 0.12s, and with 3 Saudi-Master bots
  sampling back-to-back per trick the Lua heap spikes faster than the
  incremental GC reclaims it.
  Changes (Saudi Master ISMCTS only — no UI/log-leak changes):
  • Runtime memory emergency brake in BM.PickPlay's world loop:
    collectgarbage("count") net-delta vs K.BOT_ISMCTS_MEM_BUDGET_KB
    (default 16384 — emergency-only, several× steady-state). On trip:
    stop sampling, vote with completed worlds; zero completed → the
    existing _restore(nil) heuristic fallback (unchanged). gcAvail-
    guarded so collectgarbage-less envs degrade gracefully.
  • Per-move telemetry (_lastMemKB, _lastMemCapped, _lastNumWorlds,
    _lastLegalCount, _lastTrickCount) surfaced via /baloot ismctsdiag
    (Δmem + [MEM-CAPPED] tag).
  • 3+-bot tables cap early/mid world counts more conservatively
    (trick≤2 100→50, ≤5 60→30, late 30→20) via math.min. Internal
    sample-count only — no rules/scoring/wire/UI/card-legality change.
  • Single post-move incremental collectgarbage("step", N) — never a
    full collect, never inside the per-world/per-card tight loops.
    Confirmed bounded and NOT changed: UI handPool (≤8, reused via
    clearHand), B.Log (200 fixed ring), freezeLog (200 ring, off by
    default), WHEREDNGNDB.history (200 ring, per-round).
    Tests: new test_botmaster.lua section G (38 asserts) — behavioral
    (telemetry populated + 3-bot cap = 50 @ trick0; 0-world → heuristic
    fallback) + source pins (guard structure, no full-collect literal,
    Constants tunables, ismctsdiag memory output). Full harness 1537/0;
    H1 11/0; H7 9/0.
    Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com
• docs(CLAUDE): fix CurseForge project ID (1526129 -> 1529200)
  The documented CurseForge project ID was stale — 1526129 actually
  resolves to a different addon (GOBIGnINTERRUPT). WHEREDNGN is
  project 1529200 ("Loot & Baloot", slug wheredngn). Doc-only;
  no code/packaging change.
  Co-Authored-By: Claude Opus 4.7 (1M context) noreply@anthropic.com