fix: compare .gitmem roots by real path, not string (GIT-107)

[nTEG-dev]

What

.gitmem roots are compared by real path, not by string (GIT-107).

• canonicalPath() resolves a path with fs.realpathSync.native. A path that doesn't exist yet (e.g. a destination before its first write) resolves through its nearest existing ancestor, so /var/x/.gitmem and /private/var/x/.gitmem agree even before .gitmem exists. sameDirectory(a, b) compares two paths that way.
• findStrandedProjectRoots() uses it. A candidate is also excluded when it's the same directory as an explicit GITMEM_DIR, since that's the root actually being read; before, only the home root was checked.
• migrate-root's "source and destination are the same" check uses it too.

Why

On macOS /var is a symlink to /private/var. process.cwd() reports /private/var/…, while a home or GITMEM_DIR under $TMPDIR is spelled /var/…. The string comparison saw two stores, and session_start told the user their live store wasn't being read. Reproduced on every main venue run, where the driver's store lives under $TMPDIR:

Memory store found but NOT being read
  /private/var/folders/…/gitmem-venue-baseC-main-flow-QTZRrm/.gitmem
    holds: 3 threads, 1 sessions
  gitmem reads /var/folders/…/gitmem-venue-baseC-main-flow-QTZRrm/.gitmem regardless of directory (1.8.0).

How to test

Venue (v1.8.0 setup.sql, vujndhcumgroztfkkdus; production not touched): flow now fails if session_start shows the notice, since the driver creates no stranded store. main: exit 1 ("reported the store it reads as 'NOT being read'"). This branch: exit 0.

Unit tests: 1279/1279 pass (+5, gitmem-root-symlink.test.ts):

• home reached through a symlink, cwd spelled resolved → not stranded;
• a genuinely different live store → still reported;
• GITMEM_DIR naming the project store → not stranded;
• sameDirectory with a symlink, a trailing slash, . segments and not-yet-existing leaves;
• macOS /var vs /private/var (skipped where $TMPDIR has no symlink, e.g. Linux).

4 of 5 fail on origin/main; "a different store is still reported" passes on both, as it should. No existing test encoded the old behaviour.

Checklist

• npm run test:unit passes (1279)
• New tests added (symlink, plus the macOS case)
• No schema change; verified on a v1.8.0 store
• Docs: n/a

Breaking changes

None.

🤖 Generated with Claude Code