Fix --detach-process for standalone CLI

[Travor278]

Related

• Closes --detach-process doesn't work on standalone CLI (but works fine from spawn in the SDK) #11847

What

--detach-process was copied into StartupOptions, but that field is never read by the standalone CLI, so the viewer kept running in the invoking process.

This change relaunches native-viewer CLI invocations once as a detached child:

• preserves the original CLI arguments and inserts a hidden child marker before any -- end-of-options separator to prevent recursion;
• starts a new session with setsid on Unix and uses DETACHED_PROCESS on Windows;
• disconnects standard streams so captured pipes do not keep the launcher attached;
• leaves subcommands, headless/server/web/save/test/version modes in the original process.

Focused tests cover the one-time relaunch guard, representative native-viewer invocations, rerun --detach-process -- recording.rrd, and the modes that must remain attached.

Testing

• cargo test -p rerun --lib --no-default-features --features rerun/native_viewer,rerun/run detach_ (Windows and Linux; 3 passed on each)
• cargo clippy -p rerun --lib --no-default-features --features rerun/native_viewer,rerun/run -- --deny warnings (Windows and Linux)
• cargo fmt -p rerun -- --check (Windows and Linux)
• cargo check -p rerun --lib --tests --no-default-features --features rerun/native_viewer,rerun/run (Linux)
• cargo build -p rerun-cli --no-default-features --features rerun-cli/native_viewer (Linux)
• pixi run lint-codegen
• Changed-file fast lint checks for Rustfmt, TOML formatting, repository lint, typos, and notebook stripping
• Windows lifecycle test: the launcher exited successfully in 546 ms, the detached child was observed alive, produced a screenshot, and exited without a leftover process
• Linux lifecycle test: the launcher returned in 0 seconds, the child was observed with PID == SID == PGID, wrote a 124,236-byte PNG, and terminated

Confidence

High. The dead data flow was reproduced on current main; both platform-specific paths compile, Windows and Linux process lifecycles were validated end to end, and viewer/non-viewer mode selection plus end-of-options handling have regression coverage.

Agent

🤖 This PR was opened by a coding agent. The implementation and test results were reviewed in an LLM-assisted workflow.

[wuisabel-gif]

The child marker should be inserted before any -- end-of-options separator rather than appended to the argument list. With an invocation such as rerun --detach-process -- recording.rrd, the child receives rerun --detach-process -- recording.rrd --detached-process-child. Clap treats the marker as a positional argument after --, leaving detached_process_child false, so should_relaunch_detached can recursively relaunch. Could we construct the child arguments by inserting the marker before the first --, and add a regression test for this invocation?

[Travor278]

Good catch — you were right. I was appending the marker blindly, which breaks as soon as -- is present. Fixed in d471eb3: the child marker is now inserted before the first --, and I added the exact rerun --detach-process -- recording.rrd case as a regression test. I also reran the focused tests and Clippy on both Windows and Linux. Thanks for spotting this one!