test(extraction): cover the fork's fm_agent work-directory exclusion

[Dshuishui]

The fork's own patch — fm_agent in DEFAULT_IGNORE_DIRS — has no test. It is one
line in a name set, and that set sits in a file upstream touches often, which makes
it exactly the kind of thing an upstream sync can silently drop or widen. Losing it
fails nothing: the build passes, the suite is green, the release ships, and the
index quietly grows a second copy of every function FM-Agent extracted.

This adds regression cover for it, and records the file in FORK.md's patch table —
the list a sync is checked against.

What it pins

• The exclusion applies at the project root and at any nested depth, for both
  languages the artifacts come in. Depth matters: FM-Agent's work directory sits at
  the root of the directory it was pointed at, which is not necessarily the root
  CodeGraph indexes.
• It is a whole-name match, not a prefix or substring one — fm_agent_data/,
  my_fm_agent/ and a file named fm_agent.c all stay indexed. This is what keeps
  the blast radius small.
• A .gitignore negation (!fm_agent/) takes the directory back, nested included.
  That escape hatch is what makes the any-depth match acceptable, so it belongs in a
  test rather than only in prose.
• None of it depends on git: the exclusion holds with no repository at all, and with
  the artifacts committed. The second case is not hypothetical — FM-Agent's
  incremental mode checks a commit that can contain them out into a temporary
  worktree and indexes that.

Verification

New tests 4/4; full suite 2905 passed, 0 failed.

Mutation-checked: commenting out the one line in DEFAULT_IGNORE_DIRS fails 3 of the
4 (the fourth asserts a directory is indexed, which holds without the patch too —
as expected).

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.