starwing / piecetab / 39009647941

100%

push

github

test(pt): sweep the reserve formulas past the saturated depth

The saturated fixtures stop at levels = 3 (fanout 4 needs 4^levels pieces),
so the linear terms of the piecetab budgets (`2L + 2` for remove / append /
insert, `3L + 3` for edit / splice) were only ever measured up to L = 3.
`pt_needtree` builds a right-spine tree with O(2^levels) containers: every
node on the last-slot cursor path holds PT_FANOUT children, the other
children are minimal subtrees, so an edit at the tail still has to root-push
and chain one node per level.

`remove_deep_brute` / `edit_deep_brute` sweep L = 4 and 5 across every
cursor position and three lengths each (858 cases per test): with exactly
the formula's nodes and the page allocator dead the operation must pass, and
one node short the entry reserve must refuse with PT_ERRMEM (tree and
cursor untouched).  edit spends `L + (L + 2) = 12` of its 18 at L = 5, so
the chain term does grow with L; the shape still leaves 5 nodes spare, so
the tight end of those formulas stays with the saturated witnesses.

Sensitivity: remove at `2L + 1` / `2L + 3`, edit at `3L + 2` / `3L + 4`
all fail one of the two sides.

docs: design_stitch 3.20 records the depth witness next to the saturated
one; the reserve report gets the test rows and closes the sparse-shape item.

Gates: just pt (all), remove_deep_brute / edit_deep_brute 858 cases each,
pt-cov 96.5% branches / 100% lines, clang-format --dry-run -Werror

3899 of 4029 branches covered (96.77%)

4 of 4 new or added lines in 3 files covered. (100.0%)

3966 of 3966 relevant lines covered (100.0%)

1751050.62 hits per line