fix(brain-sync): make artifact sync work on Windows (discover-new + drain)

[daveowenatl]

Automatic artifact sync (gstack-brain-sync) is fully non-functional on Windows (Git Bash): --discover-new enqueues nothing and the --once drain stages nothing, so artifacts_sync_mode=full looks active but no artifacts ever reach the repo. No error surfaces — status reports idle.

Environment

Windows 11 (26200), Git Bash (Git for Windows 2.54.0), Bun 1.3.11, Python 3.14.3 (python3 on PATH), gstack v1.43.3.0.

Root causes (three, all in bin/gstack-brain-sync)

1. discover-new path match — os.path.relpath returns backslash paths on Windows, which never match the forward-slash allowlist globs (projects/*/learnings.jsonl). Fixed by normalizing os.sep→/.

   python3 -c "import os,fnmatch;print(fnmatch.fnmatchcase(os.path.relpath(r'C:\a\projects\p\learnings.jsonl', r'C:\a'),'projects/*/learnings.jsonl'))"  # -> False
   

2. discover-new enqueue — enqueued via subprocess.run([gstack-brain-enqueue, rel]), but Windows Python can't exec a bash-shebang script, so --discover-new left the queue empty even when files matched. Now appends to the queue in-process. Same "spawnSync can't exec bash on Windows" family as fix: sourceLocalPath handles wrapped {sources:[...]} shape from gbrain v0.20+ #1571 / fix(sync-gbrain): Windows compat — capability check + brain-sync skip #1641, different code path.
3. --once drain — compute_paths_to_stage ends in print(p); Windows Python emits CRLF, the bash while IFS= read -r p keeps the trailing CR, and git add -f -- "path<CR>" fails with pathspec '…?' did not match under 2>/dev/null || true. Nothing is staged, the commit is empty, status reports idle. Now strips the trailing CR before git add.

The in-process enqueue mirrors gstack-brain-enqueue's contract rather than re-deriving it loosely: one atomic O_APPEND write per record (each line < PIPE_BUF) so a parallel writer-shim append can't interleave mid-record, and the discover cursor advances only after the write succeeds (a failed write retries instead of silently recording the file as synced). Skip-list entries are separator-normalized on both the discover and drain (compute_paths_to_stage) sides, so a backslash .brain-skip.txt entry can't be honored at discovery yet bypassed at commit — the drain is the boundary that actually stages files.

Repro

gstack-config set artifacts_sync_mode full   # ~/.gstack wired to a remote
mkdir -p ~/.gstack/projects/demo
echo '{}' > ~/.gstack/projects/demo/learnings.jsonl
gstack-brain-sync --discover-new
wc -l < ~/.gstack/.brain-queue.jsonl          # 0  (expected 1)

Testing

• test/brain-sync-windows-paths.test.ts — static invariants guarding all three causes plus the two robustness properties (atomic per-record append, normalized skip matching). Behavioral spawn tests can't run on the Windows lane (Node/Bun can't exec the bin/ shebang scripts — the same reason brain-sync.test.ts is excluded from it), so this follows the static-invariant pattern of build-script-shell-compat.test.ts. Verified red→green. Wired into windows-free-tests.yml.
• Manual end-to-end on Windows 11 / Git Bash: discover → drain → commit → push to a local bare remote lands sync: N file(s) with nested and single-level artifacts; non-allowlisted files excluded; a backslash .brain-skip.txt entry correctly keeps its file off the remote; queued records are valid compact JSONL; re-discover is idempotent.
• macOS/Linux behavior unchanged: os.sep is already / so both separator-normalizations are no-ops, there's no CRLF to strip, and the inline O_APPEND append is equivalent to the previous per-file enqueue.

Related (not in this PR)

bin/gstack-brain-restore:117 has the same relpath+fnmatch bug in its clobber check — happy to follow up separately.

🤖 Generated with Claude Code