fix: init seeds the store the server reads; repo keeps config + hooks (GIT-115)

[nTEG-dev]

GIT-115: init writes the store where the server reads it; the repo keeps config + hooks

Just upgrade the package. Nothing to change on your database. Local files only.

Implements Chris's ruling: move only the store. config.json and the hook scripts stay in the repo's .gitmem/.

What was wrong

• init seeded <cwd>/.gitmem, but the server has read ~/.gitmem since GIT-91. A fresh install found no starter lessons.
• session_start then reported the repo's .gitmem/ as "Memory store found but NOT being read", because config.json alone counted as a live store.
• The SessionStart hook read the repo's project name only with jq.
• uninstall and the telemetry commands used <cwd>/.gitmem.
• The wizard and gitmem uninstall matched hooks by the bare word "gitmem".

Fix

• bin/gitmem-root.js resolves the store the way gitmem-dir.ts does: GITMEM_DIR, then GITMEM_HOME/.gitmem, then ~/.gitmem.
• The wizard and init-scars seed learnings, threads, sessions, decisions, scar-usage, closing-payload.json and its template into that root.
• The repo's .gitmem/ holds only config.json with { project } and hooks/.
• feedback_enabled is written to the store's config.json, because the server reads it there. It is written only on opt-in.
• The stranded-store check no longer counts config.json, starter lessons or the welcome thread as evidence. A learning, thread or decision of the user's own does count now; main ignored those entirely.
• session-start.sh reads the project name with node when jq is absent.
• uninstall removes only the repo's config.json and hooks/, and leaves anything else in the repo's .gitmem/ alone. The store is shared by every project, so it is kept unless --all is given or the user declines the keep prompt.
• Wizard hook merge and uninstall hook removal use GIT-120's bin/hooks-merge.js. The file is identical to the one in fix: install/uninstall-hooks merge per event, keep foreign hooks (GIT-120) #55, so the two branches stack without a conflict.
• gitmem telemetry uses the server's root.

Evidence

Fresh-install e2e tests/e2e/git-115-fresh-install.test.ts uses a clean HOME and clean repo. It runs init, the real SessionStart hook, then the built server with session_start and recall, then uninstall.

• Branch: 9/9 pass.
• Main: 9/9 fail. Main has no ~/.gitmem/learnings.json, recall returns "No relevant scars found", session_start shows "NOT being read", and the hook gives no project hint without jq.

session_start does not list scars on any install. The test checks it two ways: session_start shows the welcome thread that init seeded into the store under the repo's project, and a recall in that session returns the starter scar "Database Migration Without Rollback Plan".

Check	Branch	Main
Unit	1362 (+5)	1357; the 6 new root-evidence cases fail
Hook suite, stock macOS PATH and slim PATH, scratch HOME	47/47 each	new project-hint test fails without jq
cli-fresh-install.test.ts	27/27	3 failing, stale since GIT-91/99
tsc	clean

Both e2e files are now wired into CI.

Old tests changed

• gitmem-root-resolution.test.ts: "counts a deliberate project-scoped install (config.json)" now asserts the opposite, per the ruling.
• gitmem-root-symlink.test.ts: the live-store fixture now holds a learning instead of config.json.
• cli-fresh-install.test.ts: every CLI call now runs with a scratch HOME. It used the developer's real HOME, so with this change init wrote into the real ~/.gitmem. The init assertions now expect the new layout. The Stop-hook wording test names its registry through GITMEM_DIR, as GIT-99 requires.

🤖 Generated with Claude Code