GuildBankLedger

v0.38.1 (2026-08-27)

Full Changelog Previous Releases

• Merge pull request #142 from RussellFeinstein/fix/overflow-pack-churn
  Keep interchangeable overflow stacks where they are (v0.38.1)
• Stamp v0.38.1
• Keep interchangeable overflow stacks where they are (#140)
  Two stacks with the same itemID and the same count are interchangeable,
  so a move between their slots has no observable result. Phase 4 ranked
  them by origSlot, which makes the target depend on where they sit now,
  and executing the plan is what changes that. A pass that ended early
  re-aimed the rest of the plan instead of shortening it: the replan came
  back larger than the work that remained, and the executor's
  non-decreasing rule stopped the sort with a residual.
  Within a run sharing both fields the assignment now minimises movement.
  A stack already inside the run's slot range stays put and only the rest
  fill the gaps. The comparator is untouched, so Phase 0's merge runs are
  unaffected, and the packing contract is unchanged because the contract
  is about which item and count occupies each slot.
  Measured on a 98-slot tab of scattered full stacks: the plan drops from
  98 ops to 74, and after executing part of it the replan goes from
  wanting 96 ops against 74 remaining to 58 against 56.
  The residual +1 or +2 is pivot overhead, not churn. A pivot parks a
  blocker and brings it back, so it costs two ops, and whether one is
  needed depends on the free-slot pattern that partial execution
  genuinely changes. The spec allows exactly that much and no more, which
  is a deliberate loosening of the first draft's "never grows": that
  invariant is unreachable while pivots exist.
• Add a failing spec for Phase 4 packing churn (#140)
  Reproduces offline what a live capture showed: executing part of a pack
  plan makes the next plan bigger, so the executor's non-decreasing rule
  stops the sort with a residual. Live it was 28 ops issued and a 124-op
  replan; here it is 24 of 98 issued and a 96-op replan.
  The fixture shape is the finding. A tab of DISTINCT items never
  reproduces it, because overflowStackOrder settles on itemID or count and
  never reaches the origSlot tiebreak. It needs many stacks sharing an
  itemID AND a count, which is what a real overflow tab looks like: dozens
  of Thalassian Phoenix Oil x20 and Light's Potential x200. Position is
  then the only thing ordering them, and executing the plan rewrites the
  positions it ordered them by.
  Three of the four specs here already pass. They are guards for the fix,
  not reproductions: a full pass still has to converge, a packed tab of
  identical stacks still has to plan zero ops, and one added stack must
  not approach a full repack.
  Also adds applyPlanPartial and redescribe next to applyPlan, since
  partial application is an ordinary state (bank closed, cancelled, pump
  stopped) rather than a synthetic one.