v0.2.28: ez create -m keeps parent fixed + full-subtree auto-restack fix

[rohoswagger]

Summary

Two changes, released as v0.2.28.

1. `ez create -m` no longer advances the parent

Previously `ez create -m` committed on the current branch before forking, advancing the parent as a side effect. Now it stashes uncommitted state, creates the branch + worktree at the parent's unchanged tip, pops the stash into the new worktree, and commits there. `-m` with `--no-worktree` is rejected (the staged changes need somewhere to live).

2. Fix stale-`parent_head` auto-restack bug

`ez commit` / `ez amend` / `ez move` only restacked direct children (`state.children_of`). After a commit on a mid-stack branch, every descendant beyond the first level was left detached — `parent_head` still recorded the old parent tip and `ez log` rendered the stack as healthy, but a later mutation on the grandchild landed on the orphaned line and lost the parent's changes until a manual `ez restack`. This violated design principle #2 ("auto-restack on every mutation… users should never have to think about rebasing").

Fix: added `StackState::descendants_topo` (full transitive subtree, parents before children) and a shared `restack::cascade_restack` helper that walks it; routed commit/amend/move through it. Because each parent is restacked before its children, a single pass converges the whole subtree. This also de-duplicates four near-identical restack loops.

Tests

• `descendants_topo_returns_whole_subtree_parents_before_children` (pure logic)
• `cascade_restack_rebases_grandchildren_not_just_direct_children` (real-git: builds main→A→B→C, commits on A, asserts both B and C are restacked and A's change reaches C)
• Plus 4 create-feature integration tests. Full suite: 226 passing, fmt + clippy clean.

🤖 Generated with Claude Code