fix: reject over-segmented skyline grids

[ivan-aksamentov]

fix/coalescent-skyline-grid-validation -> rust

Update:

• kb/decisions/coalescent-skyline-convex-inverse-tc.md: bound the every-segment-owns-a-merger guarantee to the supported segment count

• kb/issues/N-coalescent-skyline-grid-validation-incomplete.md: record that the ordering and exposure guards are now enforced

• Follow-up to: Feat/exact skyline #856

The merger-quantile boundary construction clamps interior candidates to be non-decreasing, so requesting more segments than the tree has distinct merger times produced duplicate (zero-width) or zero-exposure segments. A segment with a merger but no exposure has $M_i > 0$ and $I_i = 0$, for which the objective $I_i x - M_i \ln x$ is unbounded below; the solver masked this with a pooled warm-start fallback, so its $T_c$ was set by the smoothing prior rather than the genealogy, with no diagnostic. The recorded guarantee that every segment owns a merger held only when the requested segment count did not exceed the distinct merger times.

This rejects an over-segmented grid up front, when the requested segments exceed the distinct merger times, and validates strictly increasing boundaries and positive per-segment exposure as backstops, each with an actionable error [src]. The skyline tests move to a tree with enough lineage-count resolution for multiple segments, assert the segment-count and ordering invariants, and cover both the over-segmentation and zero-exposure errors.

Erroring is correct because a segmentation the data cannot support has no meaningful maximum-likelihood solution: the affected segment's $T_c$ is fixed by the prior, not the data. Silently capping the segment count would change the reported model without the user's knowledge, whereas a clear error lets the user choose a supported --n-skyline.

Work items

• Reject requested segments exceeding the number of distinct merger times [src]
• Validate strictly increasing boundaries and positive per-segment exposure [src]
• Rework the skyline tests to valid segment counts, asserting the invariant and both error paths [src]

Possible improvements

• Move the inline grid guards into one validated skyline-grid type (kb/issues/N-coalescent-skyline-grid-validation-incomplete.md)

[rneher]

let's hold off on that one. I was going to look into the regularization a bit more.