Severity: MEDIUM Status: ✅ RESOLVED
| Site | Defect |
|---|---|
| :376 vs :403-404 | The marker loop breaks when grid_x >= width but ref_inc = cell_rows * cell_cols counts those cells anyway → isImageReferenced never reaches 0; the slot can only be reclaimed by the forced path at :317-332 which evicts referenced images, leaving markers that name a vanished id. |
| :378-382 | Marker cells are written directly onto whatever cell is there — overwriting an existing sixel marker without decrementMainGridRef (every erase path at :736/:747/:791 does decrement). Second inflation route. |
| :527 | setupGridEviction is wired only in resize. useAltScreen swaps in a fresh Grid.init (hook null); Grid.clone (grid.zig:147-154) does not carry on_line_evict/ctx; restoreSavedGrid reinstalls a hook-less clone — history trims on those grids silently leak references. |
| :994-1002 | eraseDisplay overlap test uses hardcoded (px_height + 19) / 20 while placement uses cell_px_height (:352), and compares against img.row, which is set once at :343 and never shifted by shiftSixelAnchors (which updates anchor_row only) — after any scroll, keep/free decisions use stale geometry: visible images are freed (markers dangle) or dead ones persist. |
| :291 vs :431 | Reference check passes the full img.id but the comparison masks (img.id & 0x1FFFFF); after 2^21 images every check reports “not referenced” and slots are clobbered under live markers. |
Fix: count only placed markers; decrement before overwrite; install the eviction hook on every grid swap; compute erase overlap from anchor_row with the current cell_px_height; mask both sides of the id comparison.