Algorithm

Cascade physics

Why survivors without cleared slots below them do not move.

A correct cascade drop has a simple invariant: a symbol only moves if there are cleared slots beneath its original row. Getting this wrong is the classic cascade bug. naive implementations shift every visible cell up by the winner count, then tween back down, making untouched symbols visually jump.

pixi-reels computes per-cell fall distance from the winners list. Survivors with no winners below them are never touched. no animation job is created, their view.y stays exactly where placeSymbols put it.

The algorithm, in one column

flowchart LR
  subgraph Before["1. Before"]
    direction TB
    B0["A (row 0)"]
    B1["B winner (row 1)"]
    B2["C (row 2)"]
    B3["D winner (row 3)"]
    B4["E (row 4)"]
    B0 --- B1 --- B2 --- B3 --- B4
  end

  subgraph Remove["2. Remove winners"]
    direction TB
    R0["A"]
    R1["empty"]
    R2["C"]
    R3["empty"]
    R4["E"]
    R0 --- R1 --- R2 --- R3 --- R4
  end

  subgraph Gravity["3. Gravity"]
    direction TB
    G0["empty"]
    G1["empty"]
    G2["A · fell 2"]
    G3["C · fell 1"]
    G4["E · no move"]
    G0 --- G1 --- G2 --- G3 --- G4
  end

  subgraph Fill["4. Fill from top"]
    direction TB
    F0["X · new"]
    F1["Y · new"]
    F2["A · fell 2"]
    F3["C · fell 1"]
    F4["E · no move"]
    F0 --- F1 --- F2 --- F3 --- F4
  end

  Before --> Remove --> Gravity --> Fill

One column through the four cascade stages. Row 4 (E) never moves: no winners below it.

Stayed put · no winners below → no animation

Survivor fell · moved N slots where N = winners below

New symbol · fell from above viewport

Per-row fall distance formula

For a column with visible rows indexed 0..V-1 and winner rows winnerRows (sorted ascending), let nonWinnerRows be the surviving row indices in order. For each final row R:

Final row R	Is it a new symbol?	Starting y offset (slots above target)
`R < winnerRows.length`	yes. falls from above the viewport	R + 1
`R ≥ winnerRows.length`	no. survivor from `nonWinnerRows[R - winCount]`	R - originalRow (0 means: don't touch it)

Crucially: when offsetRows === 0, the code returns early and never creates an animation job for that cell. Its view.y is already correct from placeSymbols, and tweening it would be visually wrong.

Four canonical cases

Winner at the top

One new symbol falls in at row 0. Every survivor stays put.

Winner in the middle

Survivors above the winner each fall 1 slot. Survivors below don't move.

Winner at the bottom

Every survivor falls exactly 1 slot. New symbol fills row 0.

Two stacked winners at the top

Two new symbols fall in. No survivor moves.

API surface

reelSet.destroySymbols(winners, opts?). the "pop" primitive. Defers to each symbol's own playDestroy() so subclasses (Spine, particles) get art-appropriate disintegration (the default is a brief scale/fade implode). Lifts zIndex so destroys aren't clipped.
reelSet.refill({ winners, grid }). the "place + drop" primitive. Reads each reel's visibleRows, computes per-cell fall distances from winners, animates only the cells that actually move. Untouched survivors stay perfectly still.
reelSet.runCascade({ detectWinners, nextGrid }). the chain orchestrator. Loops detect → destroy → pause → refill until detectWinners returns []. Per-stage progress fires via cascade:chain:start / cascade:chain:end; the awaited RunCascadeResult carries the summary. Pass signal: AbortSignal for caller-driven cancellation.
computeDropOffsets(visibleRows, winnerRows). the algorithm the engine itself uses, exported so you can validate server-side gravity sims offline.

Why winners must be semantic, not diffed

Consider prevCol = [A, C, X], X at row 2 wins. The correct next col is [N, A, C]. new symbol at row 0, survivors shifted down. A naive "diff the two grids" approach reports three changed cells (row 0: A→N, row 1: C→A, row 2: X→C). Treating those as "three winners" would animate three new symbols falling from above, instead of one new + two survivors sliding by one slot. The library's API takes a Cell[] of match winners, not a diff. so the answer is unambiguous.

Your detectWinners callback in runCascade always returns the cells your game rules consider winners. The remove-symbol recipe shows the minimal case.