[matt-edmondson]

Fixes #332

What was wrong

GitCommand built git's argv from context.Remaining.Raw on the assumption that a literal -- keeps flags verbatim. It doesn't: Spectre.Console.Cli still tokenizes what follows the separator.

• -- log -1 became git log -. That's the dangerous case, since checkout - and merge - act on the previous branch.
• -- clean -fdx, -- branch -vv and -- commit -mwip were rejected with "put it after --", even though the user already had.
• -- grep -C2 foo and -- diff -U5 failed with "Short option does not have a valid name" (exit 255).

The fix

• New GitPassThrough (KtsuTools/Commands/GitPassThrough.cs). Split takes a command line whose first token is git and cuts it at the first --. Spectre gets everything before the separator, and the tail is kept as Arguments. Any other command line comes back unchanged, along with GitPassThrough.None. Any later -- stays in the tail, because that one belongs to git's own pathspec syntax.
• Program.Main now calls a new Program.Run(args, services). That method does the split, registers the GitPassThrough as a singleton, and runs the CommandApp on the shortened arguments. It is separate from Main so tests can take exactly the shipped path.
• GitCommand takes the GitPassThrough in its constructor and appends its Arguments after settings.Args and Remaining.Raw. The forgotten-separator check is unchanged, so ktools git status --short is still rejected instead of silently dropping --short.

This follows the issue's suggestion to read the real argv after the first --. Splitting before CommandApp.Run is necessary rather than optional: Spectre rejects -C2 at parse time, before the command ever runs.

Tests

All new tests are in GitCommandTests:

• TheRealAppPassesEverythingAfterTheSeparatorToGitVerbatim runs through Program.Run with a real CommandApp, as the issue's acceptance criteria ask. The cases are log -1, log -3 -p, clean -fdx, branch -vv, grep -C2 foo, diff -U5 and log '--format=%an %s' -1, and each checks the exact argument string git receives.
• TheRealAppKeepsPositionalArgumentsBeforeTheSeparator covers git log --path … -- -1, which should give log -1.
• TheRealAppStillRejectsAFlagWrittenWithoutTheSeparator covers git status --short, which should exit 1 and run nothing.
• Three Split unit tests cover a later -- staying in the tail, a non-git command being left alone, and a git command line with no separator being left alone.

With Split turned into a no-op, so that Spectre sees the whole command line as it does on main, all 8 separator cases fail. The rejection test passes either way. The full suite passes: 397/397.

CI note

On main, restore currently fails with NU1903: SixLabors.ImageSharp 3.1.12 has a known high-severity advisory. That failure is unrelated to this change, and it is the same one #334, #337 and #338 report. To run the build and tests above locally, I passed -p:WarningsNotAsErrors=NU1903.

---

Generated by Claude Code

[matt-edmondson]

The ci / .NET / Test on … checks fail at restore with error NU1903: SixLabors.ImageSharp 3.1.12 has a known high-severity vulnerability (GHSA-v76p-62qx-wwq2).

The same checks are red on main at 3382f1d, and this PR doesn't touch the ImageSharp reference, so the failure isn't caused by this PR.

I don't know of a fix I could bring into this PR. #270, the bump to 3.2.0, is red as well because 3.2.0 needs a Six Labors license in CI (see ktsu-dev/IconHelper#155).

Locally, with -p:WarningsNotAsErrors=NU1903, the full suite passes: 397/397.

---

Generated by Claude Code