(bug fix): prune usage after change promotion in loop usage finalization

[eytan-starkware]

A loop/closure that reads an ancestor as a snapshot (@a.b) while mutating a
nested field (a.b.c) silently dropped the mutation, returning a stale value
(e.g. 1 instead of 4).

In Usage::finalize_as_scope, when a changed child has an ancestor that is
read as a snapshot, the ancestor is promoted into usage/changes so the loop
reads and writes back the whole ancestor. The usage-pruning pass ran before
the promotion, so the promoted ancestor's descendants survived in usage,
violating the antichain invariant. Downstream the loop signature then listed
both the ancestor and its descendants, and the reconstruction of the returned
ancestor overwrote the mutated child with the stale snapshot input.

Fix: reorder the finalization passes — snap_usage, then changes (promotion),
then usage — so the usage pass prunes descendants subsumed by any promoted
ancestor. Adds an e2e sanity repro.

[reviewable-StarkWare]

This change is 

[eytan-starkware]

• (bug fix): prune usage after change promotion in loop usage finalization #10201 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[cursor]

PR Summary

Medium Risk
Touches core loop/closure variable-usage finalization in the semantic layer, which drives loop signatures and write-back; wrong ordering could affect many mutating loops, though scope is a focused reorder plus one regression test.

Overview
Fixes a semantic bug where loops that snapshot an ancestor (@a.b) while mutating a nested field (a.b.c) could return a stale value because mutations were effectively dropped at codegen.

In Usage::finalize_as_scope, the usage-pruning pass now runs after snap_usage pruning and the changes pass (which promotes a snapshotted ancestor into usage/changes). That way, when promotion happens, descendant member paths are pruned from usage, restoring the antichain invariant and ensuring the loop reads/writes the whole promoted ancestor instead of mixing ancestor + child paths that overwrite the mutation.

Comments in the promotion logic are updated to describe ancestor promotion vs. stale child snapshots. An e2e CASM sanity case loop_snap_mut (input 2 → output 4) locks in the behavior.

^{Reviewed by Cursor Bugbot for commit 9f8cd38. Bugbot is set up for automated code reviews on this repo. Configure here.}

[TomerStarkware]

@TomerStarkware reviewed 2 files and all commit messages, and made 1 comment.
Reviewable status: complete! all files reviewed, all discussions resolved (waiting on orizi).