Restore and harden pager suppression for wait-mode commands (APP-5310)

[warp-agent-staging]

Description

The agent stopped reliably suppressing pagers for gh/VCS commands. Root cause: ShellCommandExecutor::execute only decorates pager-prone commands (wrapping them so their output is piped through cat, preventing the pager from activating) when uses_pager && wait_until_completion. The server was unconditionally reporting wait_until_completion = false for the modern run_shell_command tool call, in both wait and interact modes — a bug now fixed in the companion PR warpdotdev/warp-server#14303. With that field always false, the client's decoration gate never fired even when the model correctly set uses_pager: true, letting commands like gh pr view or git log drop into the user's pager and hang or truncate.

This PR restores the client's uses_pager && wait_until_completion gate (now that the server reports the field correctly again — true for wait mode, false for interact mode), and hardens the decoration itself:

• Mode-aware gating: only wait-style commands are decorated. interact-mode commands (REPLs, dev servers, TUIs, or a pager the subagent means to page through itself) are left alone so the subagent keeps live PTY control.
• Exit-status preservation: piping through cat used to mask the wrapped command's real exit status (cat being the last stage of the pipe always exits 0). The decoration now re-asserts the original exit status via PIPESTATUS/pipestatus (Bash/Zsh) or a nested fish -c process (Fish).
• Heredoc safety (CSAT-10167 class): the closing group syntax (), end) now lands on its own line, so a command ending in a bare heredoc terminator (e.g. EOF) isn't corrupted by trailing characters on that line.

Linked Issue

Linear: APP-5310

Testing

• execute_decorates_pager_command_when_waiting_for_completion / execute_does_not_decorate_pager_command_in_interact_mode: prove the wait-vs-interact gate.
• turn_off_pager_for_command_preserves_nonzero_exit_status_in_bash: spawns a real bash subprocess to prove a failing command's exit code survives decoration.
• turn_off_pager_for_command_preserves_heredoc_in_bash / turn_off_pager_for_command_handles_command_without_trailing_newline_in_bash: spawn real bash subprocesses to prove heredoc-ending and no-trailing-newline commands aren't corrupted.
• turn_off_pager_for_command_zsh_uses_pipestatus_and_safe_closer, ..._fish_uses_nested_process_for_exit_status, ..._powershell_closer_on_own_line: structural checks for the other shells (not run as real subprocesses, since Zsh/Fish/PowerShell aren't guaranteed present in CI).

Ran to completion locally:

• cargo fmt -- --check
• cargo clippy -p warp --tests -- -D warnings
• cargo test -p warp --lib shell_command (11/11 passing)

On visual proof: I don't have computer-use tooling available in this environment to capture a screen recording of Agent Mode. The tests above exercise the real, observable behavior end-to-end (actual shell processes, real exit codes, real heredoc parsing), which is the strongest evidence I can produce here. If a recording is still required before merge, it needs a human or a computer-use-capable agent to capture it.

• I have manually tested my changes locally with ./script/run

Agent Mode

• Warp Agent Mode - This PR was created via Warp's AI Agent Mode

Conversation: https://staging.warp.dev/conversation/cbb2de15-6edb-4bc9-851e-3876c7b523f3
Run: https://oz.staging.warp.dev/runs/019ff1c9-37d2-7b36-bc92-33f392a1c482

This PR was generated with Oz.