feat(extraction): exclude FM-Agent's work directory by default

FORK.md

@@ -3,14 +3,26 @@
 FM-Agent's maintenance fork of
 [colbymchenry/codegraph](https://github.com/colbymchenry/codegraph).
 
-**Pinned base:** upstream `v1.5.0`. Upstream shipped the C macro-attribute
-extraction fix (issue #1211, PR #1311) in v1.5.0, so this base carries it
-natively — the fork no longer needs its own patch on top.
+**Pinned base:** upstream `main` at `c6aaa20358cd6adcd04b87bdef8e5803ad146f3a`
+(2026-08-07). That is *past* `v1.5.0` and before whatever upstream tags next:
+`v1.5.0` was 104 commits behind, and the incremental-sync convergence work, the
+first-class `union` nodes and the WAL growth fix in between all matter to how
+FM-Agent reads the graph. The version marker keeps saying `1.5.0-fmagent.N`
+because it is the last upstream *release* this descends from — the exact base is
+this commit.
 
-**Patches:** none currently — this base is byte-identical to upstream v1.5.0. The
-fork is kept as standing infrastructure so an urgent fix upstream hasn't shipped
-can be applied and released quickly if needed; each such fix would live as a
-merged pull request here.
+Upstream shipped the C macro-attribute extraction fix (issue #1211, PR #1311) in
+v1.5.0, so the base carries it natively; the fork no longer needs its own patch
+for it.
+
+**Patches:** one, listed below. Apart from it the tree matches the pinned base, so
+the fork stays cheap to re-sync. Every patch lives as a merged pull request here
+and must be **re-applied on each upstream sync** — if a merge drops one, this list
+is what catches it.
+
+| File | Patch |
+|------|-------|
+| `src/extraction/index.ts` | `fm_agent` added to `DEFAULT_IGNORE_DIRS`. FM-Agent writes its work directory into the project it analyses, holding one copy of every function it extracts plus the scripts staged to produce them, so indexing it lists each function twice and mixes tool code in with project code. Upstream deliberately keeps names that could be real source out of that list, so this stays fork-only; a project that does own an `fm_agent/` directory opts back in with a `.gitignore` negation (`!fm_agent/`). |
 
 **Version marker:** `codegraph --version` → `1.5.0-fmagent.N` identifies a build
 from this fork. Note this is a SemVer pre-release of `1.5.0`, so it sorts *below*
@@ -27,8 +39,11 @@ bundles (darwin/linux/windows, arm64/x64) with the native Rust extraction
 kernel. FM-Agent's `install.sh` pins one via `CODEGRAPH_VERSION`.
 
 **Policy:** pin to a base, don't chase upstream — update only when we need
-something a newer upstream release carries. Base updates land via a pull request
-that **merges** the target upstream tag into `main` (so upstream stays an ancestor
-— blame, audits and future syncs follow upstream history) and re-applies this
-fork layer; merge such PRs with a merge commit, not a squash. Upstream is tracked
-via the `upstream` git remote.
+something a newer upstream state carries. Prefer a tagged release; an untagged
+`main` commit is fair game when what we need is not released yet, and then the
+commit id *is* the base (record it above, and verify the build ourselves —
+nothing upstream has vetted it). Base updates land via a pull request that
+**merges** that upstream point into `main` (so upstream stays an ancestor — blame,
+audits and future syncs follow upstream history) and re-applies this fork layer;
+merge such PRs with a merge commit, not a squash. Upstream is tracked via the
+`upstream` git remote.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@colbymchenry/codegraph",
-  "version": "1.5.0-fmagent.1",
+  "version": "1.5.0-fmagent.2",
   "description": "Supercharge AI coding agents with semantic code intelligence — surgical context, fewer tool calls, faster answers. 100% local.",
   "repository": {
     "type": "git",

src/extraction/index.ts

@@ -196,6 +196,14 @@ const DEFAULT_IGNORE_DIRS: ReadonlySet<string> = new Set([
   '__history', '__recovery',
   // Generic cache
   '.cache',
+  // FORK-SPECIFIC (see FORK.md): FM-Agent's work directory. It holds one copy of
+  // every function FM-Agent extracts from the project, plus the scripts it stages
+  // to produce them, so indexing it lists each function twice and mixes tool code
+  // in with project code. Unlike the entries above, this name is not from the
+  // github/gitignore templates and could plausibly be real source — which is why
+  // it stays fork-only. A project that does own an `fm_agent/` directory opts back
+  // in with a `.gitignore` negation (`!fm_agent/`), as for any default-ignored dir.
+  'fm_agent',
 ]);
 
 /**