[Tryanks]

Summary

Collapsing a Work Log can leave the following assistant answer at the timeline's left edge, with its timing footer over the last line. Keep each virtualized chat row identified by its existing content identity.

 ChatView timeline row
-  anonymous root (gpui-fast matches it by list index)
+  .id(("timeline-row", row.identity))

Expanding a Work Log inserts virtualized activity rows, changing the following message's list index. gpui-fast 0.1.8 can reuse its retained Markdown layout under the changed ancestor path without measuring it again, losing ancestor offsets. The carry belongs at the chat list owner, using the identity already owned by TimelineRow; it adds no state, helper, dependency, or test-only production seam.

Checked the other collapsibles: command output, tool details, thinking blocks, disclosures, and inline diffs append their bodies as final children. They do not omit a child before a following sibling when toggled. Changed-file chips and their more/fewer control already have element IDs. The work-log header renders separately from its virtualized activity rows, so an empty body placeholder would not address this case.

Evidence

• Before: chat::tests::collapsing_work_log_preserves_following_message_bounds fails alone and with the chat suite (34 other tests pass). The 720px-wide answer moves from (340, 640) to (20, 736); its last paragraph ends at y=830 while the footer remains at y=770. Removing only the row ID makes the final regression test fail again for the same bounds mismatch.
  After: The real ChatView is seeded through the existing host fixture, rendered, expanded and collapsed through simulated clicks, then drawn. The answer's first/last paragraph bounds and footer bounds equal their initial values, and the footer clears the final paragraph, in light and dark themes at 1400px and 393px. The chat suite passes all 35 tests.
• The failing assertion protects the reported visible geometry regression; existing virtualization and manual-collapse tests did not assert the following answer's bounds. CONTRIBUTING's failure review settles this at step 3: a real product failure, fixed at its owner. No existing tests were rewritten, loosened, or removed. chat/mod.rs: 34 → 35 tests; +4 production lines, +76 test lines.
• Validation (all cargo subcommands used /Users/tryanks/.local/share/mise/installs/mr-boxington/latest/.mise-bins/mbx):
  • mbx fmt --all --check and git diff --check — passed.
  • mbx clippy --workspace --all-targets --locked -- -D warnings — passed.
  • mbx nextest run --workspace --locked --test-threads 4 --no-fail-fast — 1,242 passed, 13 skipped; no leaks.
  • RUSTFLAGS='-D warnings' mbx check -p tcode-web --target wasm32-unknown-unknown --locked — passed.
  • mbx machete — passed; no unused dependencies.
• Per the maintainer's updated instructions, no app was launched or driven, no new screenshots were taken, and nothing was installed. Desktop visual verification is left to the maintainer; the two-theme/two-width checks above are suite geometry checks. No upstream files were changed or upstream actions taken.

Merge Danger

Door: two-way

Blast Radius: chat

Changes retained element identity for chat timeline rows; existing layout and scroll tests run against it. No protocol, persistence, or dependency changes.