fix(penpal): eliminate unnecessary git check-ignore processes and full project walks

apps/penpal/ERD.md

@@ -98,10 +98,10 @@ see-also:
 
 ## Cache
 
-- <a id="E-PENPAL-CACHE"></a>**E-PENPAL-CACHE**: An in-memory cache (`sync.RWMutex`-protected) holds the full project list and per-project file lists. `RefreshProject()` walks the filesystem; `RefreshAllProjects()` runs in parallel with no concurrency limit. `RescanWith()` replaces the project list while preserving git enrichment.
+- <a id="E-PENPAL-CACHE"></a>**E-PENPAL-CACHE**: An in-memory cache (`sync.RWMutex`-protected) holds the full project list and per-project file lists. `RefreshProject()` walks the filesystem for full rescans; `RefreshAllProjects()` runs in parallel with a concurrency limit of 4. `RescanWith()` replaces the project list while preserving git enrichment and cached file data for unchanged projects — only new or source-changed projects are rescanned. Incremental mutations (`UpsertFile`, `RemoveFile`) update individual cache entries without walking the filesystem.
   ← [P-PENPAL-PROJECT-FILE-TREE](PRODUCT.md#P-PENPAL-PROJECT-FILE-TREE)
 
-- <a id="E-PENPAL-SCAN"></a>**E-PENPAL-SCAN**: `scanProjectSources()` walks `RootPath` recursively for tree sources, skipping `.git`-file directories (nested worktrees), gitignored directories (via `git check-ignore`), source-type `SkipDirs`, and non-`.md` files. Gitignore checking is initialized once per scan via `newGitIgnoreChecker(projectPath)`, which detects whether the project is a git repo; non-git projects skip the gitignore check gracefully. On write or read failure (partial 4-field response), the checker disables itself (`isGitRepo=false`) to prevent permanent stream desync. The source's own `rootPath` is never checked against gitignore (the `path != rootPath` guard ensures registered sources always scan). Files returning `""` from `ClassifyFile()` are hidden. Files are de-duplicated by project-relative path (first source wins) and sorted by `ModTime` descending. `EnsureProjectScanned()` is the lazy-scan entry point — it uses write-lock gating (`projectScanned` set under `mu.Lock` before scanning) to prevent concurrent requests from triggering duplicate filesystem walks. `projectHasAnyMarkdown()` performs a cheap startup check that aligns with the full scan: it uses the same gitignore checking, skips `.git`, `node_modules`, `.hg`, `.svn`, and nested worktree directories, and stops at the first `.md` file found.
+- <a id="E-PENPAL-SCAN"></a>**E-PENPAL-SCAN**: `scanProjectSources()` walks `RootPath` recursively for tree sources, skipping `.git`-file directories (nested worktrees), gitignored directories (via `git check-ignore`), source-type `SkipDirs`, and non-`.md` files. Gitignore checking is initialized once per scan via `newGitIgnoreChecker(projectPath)`, which detects whether the project is a git repo; non-git projects skip the gitignore check gracefully. On write or read failure (partial 4-field response), the checker disables itself (`isGitRepo=false`) to prevent permanent stream desync. The source's own `rootPath` is never checked against gitignore (the `path != rootPath` guard ensures registered sources always scan). Files returning `""` from `ClassifyFile()` are hidden. Files are de-duplicated by project-relative path (first source wins) and sorted by `ModTime` descending. `EnsureProjectScanned()` is the lazy-scan entry point — it uses write-lock gating (`projectScanned` set under `mu.Lock` before scanning) to prevent concurrent requests from triggering duplicate filesystem walks. `projectHasAnyMarkdown()` performs a cheap startup check that aligns with the full scan: it uses the same gitignore checking, skips `.git`, `node_modules`, `.hg`, `.svn`, and nested worktree directories, and stops at the first `.md` file found. `CheckAllProjectsHasFiles()` runs with a concurrency limit of 4 to cap subprocess spawning. `ResolveFileInfo()` resolves source membership for a single absolute path without spawning a git check-ignore process — it applies the same source-priority, SkipDirs, RequireSibling, and ClassifyFile rules as the full walk.
   ← [P-PENPAL-PROJECT-FILE-TREE](PRODUCT.md#P-PENPAL-PROJECT-FILE-TREE), [P-PENPAL-FILE-TYPES](PRODUCT.md#P-PENPAL-FILE-TYPES), [P-PENPAL-SRC-DEDUP](PRODUCT.md#P-PENPAL-SRC-DEDUP), [P-PENPAL-SRC-GITIGNORE](PRODUCT.md#P-PENPAL-SRC-GITIGNORE)
 
 - <a id="E-PENPAL-TITLE-EXTRACT"></a>**E-PENPAL-TITLE-EXTRACT**: `EnrichTitles()` reads the first 20 lines of each file to extract H1 headings. Titles are cached and shown as the primary display name when present.
@@ -239,7 +239,7 @@ see-also:
 - <a id="E-PENPAL-SSE"></a>**E-PENPAL-SSE**: `GET /events` is a long-lived SSE stream using `event: change` messages. Event types: `projects`, `files`, `comments`, `agents`, `navigate`. Each event carries optional `project`, `path`, `worktree` fields.
   ← [P-PENPAL-REALTIME](PRODUCT.md#P-PENPAL-REALTIME)
 
-- <a id="E-PENPAL-WATCHER"></a>**E-PENPAL-WATCHER**: The file watcher bridges `fsnotify` to SSE. Two-tier watch strategy: base (shallow workspace + project root directories) and dynamic (deep, per-focus). Debounce at 100ms per event key.
+- <a id="E-PENPAL-WATCHER"></a>**E-PENPAL-WATCHER**: The file watcher bridges `fsnotify` to SSE. Two-tier watch strategy: base (shallow workspace + project root directories) and dynamic (deep, per-focus). Debounce at 100ms per event key. File events use an accumulating debounce that collects per-file paths and ops during the debounce window, then applies incremental cache mutations (`UpsertFile`/`RemoveFile`) instead of full project walks. Only `.md` file events trigger cache updates — non-`.md` Create events (e.g., scanner temp files) are filtered out. Structural events (workspace changes, source auto-detect) still use the original debounce with full discovery.
   ← [P-PENPAL-REALTIME](PRODUCT.md#P-PENPAL-REALTIME), [P-PENPAL-LIVE-UPDATE](PRODUCT.md#P-PENPAL-LIVE-UPDATE)
 
 - <a id="E-PENPAL-FOCUS"></a>**E-PENPAL-FOCUS**: `windowFocuses map[string]focusTarget` (one entry per browser window) drives dynamic watches. Union of all window focuses determines the watched set. File focus watches only the file's parent directory. Project focus watches all source directories + `.penpal/comments/`.

apps/penpal/TESTING.md

@@ -64,7 +64,7 @@ see-also:
 | Source Types — anchors (P-PENPAL-SRC-ANCHORS, SRC-ANCHORS-GROUP, SRC-ANCHORS-NESTED) | discovery_test.go (TestClassifyAnchorsFile, TestGroupAnchorsPaths, TestGroupAnchorsPaths_MarkerOnlyModule, TestAnchorsFileOrder, TestAnchorsRequireSibling) | — | — | — |
 | Source Types — claude-plans (P-PENPAL-SRC-CLAUDE-PLANS) | — | — | — | — |
 | Source Types — manual (P-PENPAL-SRC-MANUAL) | — | — | grouping_test.go (TestBuildFileGroups_ManualSourceDirHeadings) | — |
-| Cache & File Scanning (E-PENPAL-CACHE, SCAN) | cache_test.go (TestCheckAllProjectsHasFiles, TestProjectHasAnyMarkdown_SkipsGitignored, TestProjectHasAnyMarkdown_SkipsVCSDirs, TestAllFiles_DeduplicatesAllMarkdown, TestEnsureProjectScanned_NoDuplicateScans) | — | — | — |
+| Cache & File Scanning (E-PENPAL-CACHE, SCAN) | cache_test.go (TestCheckAllProjectsHasFiles, TestProjectHasAnyMarkdown_IgnoresGitignore, TestProjectHasAnyMarkdown_SkipsVCSDirs, TestAllFiles_DeduplicatesAllMarkdown, TestEnsureProjectScanned_NoDuplicateScans, TestResolveFileInfo, TestUpsertFile, TestRemoveFile, TestRescanWith_PreservesUnchangedProjects, TestSourcesChanged) | — | — | — |
 | Worktree Support (P-PENPAL-WORKTREE) | discovery/worktree_test.go, cache/worktree_test.go | Layout.test.tsx | worktree_test.go (API + MCP) | — |
 | Worktree Dropdown (P-PENPAL-PROJECT-WORKTREE-DROPDOWN) | — | Layout.test.tsx | — | — |
 | Git Integration (P-PENPAL-GIT-INFO) | — | — | — | — |