Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@internal/mcpserver/mcpserver.go`:
- Around line 569-572: In the loop that builds the project list (the for _, p :=
range projects block) replace the sb.WriteString(fmt.Sprintf(...)) pattern with
a direct fmt.Fprintf call to write to sb (fmt.Fprintf(sb, "- **%s** (id: `%s`,
path: `%s`) — %d memories\n", p.Name, p.ID, p.Path, count)). Use sb as the
io.Writer and handle/ignore the returned (int, error) per project (e.g., assign
to _, _ or check and log) to satisfy staticcheck QF1012 and avoid the
unnecessary intermediate allocation; keep the call site referencing sb and
s.store.CountMemories unchanged.

---

Nitpick comments:
In `@internal/memory/store.go`:
- Around line 117-121: The SQL heuristic `path NOT LIKE '/%'` in the
QueryRowContext call incorrectly treats Windows absolute paths as non-absolute;
update the logic to be platform-agnostic by recording path metadata instead of
relying on string patterns: add a boolean column (e.g., projects.is_mcp_created
or projects.is_absolute) and set it when inserting/updating projects, then
change the query in the function that calls s.db.QueryRowContext (the query that
scans into dupID) to filter on that boolean flag instead of `path NOT LIKE
'/%'`; ensure all insert/update code paths that create project rows (where name,
id, path are set) also set the new flag.

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@internal/claudeimport/import_test.go`:
- Around line 289-293: Replace the unchecked defer db.Close() with a checked
cleanup: call t.Cleanup(func() { if err := db.Close(); err != nil {
t.Fatalf("closing memory DB: %v", err) } }) so any error returned by
memory.OpenDB's Close is reported; update the block where
memory.OpenDB(":memory:") is called and remove the plain defer db.Close().

In `@internal/claudeimport/import.go`:
- Around line 46-60: encodeProjectPath and ClaudeMemoryDir fail for native
Windows paths because backslashes and drive letters make the joined path rooted;
normalize the projectPath in ClaudeMemoryDir (or inside encodeProjectPath) by
calling filepath.Clean and filepath.ToSlash, strip any Windows volume name (use
filepath.VolumeName to detect and strings.TrimPrefix to remove it), then replace
remaining "/" with "-" (and TrimLeft any leading "-" so a leading slash or
removed drive doesn't produce a leading dash). Update the encodeProjectPath
and/or ClaudeMemoryDir flow to operate on the normalized path so C:\work\ghost
becomes "C-work-ghost" or "work-ghost" consistently and the final
filepath.Join(home, ".claude", "projects", encoded, "memory") is not treated as
rooted on Windows.
- Around line 138-149: parseFrontmatter currently only recognizes LF delimiters
("---\n") so CRLF frontmatter ("---\r\n") is treated as body; update
parseFrontmatter to accept both LF and CRLF boundaries by checking for both
starting delimiters (e.g., "---\n" and "---\r\n") and searching for matching
closing delimiters ("\n---" and "\r\n---") when computing end and slicing body;
ensure fm extraction and body = raw[...] calculations account for the chosen
delimiter length (2 vs 4 chars) so indices remain correct and the returned
name/description/fileType parsing still uses the cleaned fm.

In `@internal/mcpserver/mcpserver.go`:
- Around line 271-277: The code swallows the error returned by
claudeimport.Import which hides import failures; update the clause in
mcpserver.go (around CountMemories/ListProjects logic) to capture the returned
error from claudeimport.Import(ctx, s.store, args.ProjectID, p.Path, s.logger)
and handle it instead of discarding it — at minimum log it via s.logger (or
return it to the caller) so the errcheck violation is resolved and first-contact
import failures are visible; ensure you modify the block that calls
CountMemories, ListProjects and claudeimport.Import to check the import error
and act accordingly.
- Around line 86-87: ghost_memory_save is resolving name-based args.ProjectID to
the hash ID then calling EnsureProject(path=args.ProjectID), which overwrites
the existing projects.path with the hash; update ghost_memory_save so that if
the project already exists (lookup via the resolved ID or EnsureProject result
indicates existing), you do not replace the stored Path/Name—either skip the
EnsureProject call or call EnsureProject with the original filesystem path/name
preserved and avoid updating projects.path when a row already exists; ensure
this check uses the same identifiers used elsewhere (args.ProjectID, the
resolved project hash, and the EnsureProject return) to detect existing projects
and preserve projects.path so ghost_list_projects and import/merge gates keep
working.

In `@internal/orchestrator/orchestrator.go`:
- Around line 84-96: The goroutine uses the initial CountMemories result to
decide onboarding, which can race with other imports; after running
claudeimport.Import (or if importErr != nil), re-query the store by calling
o.store.CountMemories(ctx, projCtx.ID) inside the goroutine and only call
onboardProject(bgCtx, o.client, o.store, projCtx, o.logger) if that fresh count
is still 0 and o.client != nil; also handle and log any count error (treat
non-nil error as conservatively non-zero to avoid onboarding duplication).