Severity: LOW
Status: Fixed
None observed in production. Both defects are latent landmines that only fire
if some other bug first corrupts the history ring bookkeeping
(history_count > 0 or history_start > 0 while history.items.len == 0).
If triggered, the server would crash (index-out-of-bounds panic in Debug/
Safe builds) or read out of bounds (ReleaseFast), taking down every attached
session — the failure mode bug #380 explicitly tried to avoid.
Two reachable-by-corruption sites were found in code review of the bug #443 fix:
getHistoryLine empty-history branch: returned &self.history.items[0]
when history.items.len == 0 — an out-of-bounds read of an empty list.scrollUp reuse-path fallback: insert_idx = if (items.len > 0) ... else 0
still executed self.history.items[insert_idx] afterwards; with an empty
list the else 0 fallback indexes slot 0 out of bounds. The same fallback
pattern exists in clearHistory and setHistoryLimit eviction loops.getHistoryLineMut: zero callers in the tree, carried the identical
broken empty-list guard, and cannot be hardened with a shared read-only
sentinel (a mutating return must not alias a global other grids could
observe or deinit).Bug #443 standardized ring arithmetic on % history.items.len and added
items.len > 0 guards to survive a corrupt ring, but the guard fallbacks chose
index 0 — which is only safe if the list is non-empty, the exact condition
being guarded against. A corrupted state therefore walked straight into the
out-of-bounds access the guard was meant to prevent. getHistoryLine’s
fallback had the same flaw behind a “callers always check idx < historyLen”
assumption.
src/grid.zig:
added a container-level const EMPTY_HISTORY_LINE: GridLine = .{} and made
getHistoryLine return &EMPTY_HISTORY_LINE for an empty ring. All callers
already bounds-check (x < line.cells.items.len, .wrapped), so a blank
line is invisible to rendering and selection. The sentinel is never
allocated and never owned, so it cannot leak or be double-freed.normalizeHistoryRing(), which resets history_count/history_start to 0
when history.items.len == 0, called at the top of scrollUp,
clearHistory, and setHistoryLimit. After normalization, scrollUp
provably takes the append path (correct for an empty ring), so the else 0
fallbacks are dead code; they are kept as belt-and-braces defense.getHistoryLineMut (zero callers)."Grid recovers from corrupt empty history ring without OOB
access (bug #444)" corrupts the ring state, exercises the read path,
scrollDown, clearHistory, setHistoryLimit, and scrollUp, and asserts
blank reads plus self-healed bookkeeping.