# CodeRabbit follow-up verification

PR [#24352](https://github.com/pingdotgg/t3code/pull/24352), head [9cb8624473](https://git.cafe/pingdotgg/t3code/commit/9cb86244731908f47ab3753fbdcab94391a1fb71).

Both findings were reproduced and fixed:

1. A same-viewer resumed attach overwrote the grid from a later resize. New Manager regression failed before the fix with expected100columns/actual54. Attach now only initializes a missing grid; resize owns subsequent updates. The regression also verifies both phone-stream releases restore130columns and only one PTY was spawned.
2. Resolving the supplied seed home while comparing only the lexical live-home prefix missed symlinked ~/.t3 targets. Scratch-HOME reproduction wrote the sentinel workspace and an empty SQLite database through that alias before failing on missing tables. The guard now compares lexical and resolved inputs/live homes and their descendants before any write. All target/child/alias/relative/.. live variants are rejected; isolated and isolated-symlink homes remain accepted. No real live home was written.

Validation: Manager+OutputProtocol61/61PASS; scopedlintPASS withoneexistingprefer-array-findwarning;formatPASS;gitdiff--checkPASS;Python3.12parsePASS;scratchguardchecksPASS. No renderer/native/contract/TUI-template changes. The full earlier gate and captured792 evidence retain their exact historicalrevisionlabels.

Complexity: viewer initialization O(1). Seed guard performs four path-ancestry checks, linear in path length; no regex. A3000-token malformed CLI invocation was rejected with exit2 in58.868ms and no scratch writes. This is malformed-argument handling, not a native/parser performance guarantee.

No fourth fullreview was opened after the bounded3rounds. Android/Electron and fullmatrixgaps remainexplicit. No merge.
