fix(server): setup scripts complete with shell line editors#17817

Worktree setup can remain on “Starting setup script” when a Bash line editor treats the completion wrapper’s carriage returns as separate submissions. The incomplete block never prints its completion marker, so the waiting agent does not start.

Submit complete POSIX commands, decode the script with the shell’s printf, and evaluate it in a subshell before reporting its exit status. Short scripts use one terminal line; long payloads use bounded, independently complete assignments before evaluation so input cannot be truncated by the terminal’s line limit during shell startup. A suffix preserves trailing newlines through command substitution. This retains heredocs, literal text, stdin reads, explicit exits, and completion after syntax errors. Fish and PowerShell retain their existing behavior.

Maintainer takeover of #17817, preserving @sgaabdu4’s commits and credit for finding the line-editor failure and trailing-newline edge case. Current main was merged into the contributor branch, then the implementation and tests were rebuilt. This is one focused server bug fix under the contribution guide’s small-fix exception; the takeover claim records maintainer ownership.

Verification:

• The existing test file now has seven tests, down from the original PR’s nine: the existing test plus six focused behavior cases. The initial six-test suite fails twice on unfixed main. An independent review found that a single long input line could be truncated before readline starts; the added regression fails on that revision. The final implementation passes all seven. Disabling only the newline-preserving suffix produces the expected trailing-backslash output regression.
• Changed-file lint and formatting, git diff --check, and server-only typecheck pass.
• Live Linux PTY proof uses the actual setup-runner service with a real node-pty transport and Bash 5.2.21. The final head passes all six cases with both plain readline and ble.sh nightly, writing immediately after shell spawn. The long case uses 12 complete lines, at most 1,023 UTF-8 bytes each including Enter. Loading the main version of the runner with ble.sh leaves the original five cases waiting in multiline input without completion; the earlier one-line rebuild also hangs on the long case before this correction.

This is a server and terminal-transport change, so the proof is shell output rather than screenshots. PTY processes used for verification were stopped. Browser rendering, Bash 5.3 specifically, native zsh/fish/PowerShell, macOS, Windows, iOS, Android, Electron, and relay/tunnel connections were not exercised on this Linux machine.

Known cosmetic limitation: ble.sh can redraw a partial internal assignment into setup progress for a long script. The recorded output and completion remain correct; this change keeps the existing echo filter.

Independent whole-diff review: round one found the long-line truncation regression, which is now fixed and covered by a focused regression and fresh live proof. The final Opus/Sol round found no blockers after the fixes and convention cleanup. The test harness uses scoped Effect subprocesses; a character loop encodes shell input without a lint override. No new diagnostic or ESLint suppression is introduced.

Tests: Codex gpt-6-astra (high). Implementation: Codex gpt-6.1-sol (high). Reviews: Claude claude-opus-5-5 and Codex gpt-6-sol. Harnesses: Codex and Claude Code through T3 Code.