fix: session_close honours or refuses an explicit session_id; gitmem_session_id alias (GIT-122)

[nTEG-dev]

What

Fixes GIT-122: session_close must honour or refuse an explicit session id. Root cause (GIT-122, 19:07Z comment): on 2026-10-04 the Claude remote-devices bridge strips a parameter named session_id, so gitmem received no id, silently fell back to its registry-bound session, and overwrote an already-closed session (ed6b400e). Server log at 13:39Z: Recovered session_id from registry: ed6b400e, a path that only runs when no id arrives.

Stacked on #68 (GIT-123, the e2e sandbox fix; that commit is included here). Once #68 merges I rebase and this diff is the two GIT-122 commits only.

Changes (src/tools/session-close.ts, schema/types/tool definition)

1. gitmem_session_id alias: same meaning as session_id, under a name a proxy won't claim. Both given and different → refused. The tool definition now requires only close_type.
2. Invariant: the requested id is captured at entry; a resolved id that differs is refused before thread sync or any write, and again just before the upsert (asserting on the row id, not a resolved promise). Retroactive closes (new row by design) are exempt.
3. Re-close: refused unless reclose: true. A close that only partly persisted (threads not synced) is flagged partial_persist and stays retryable. Free-tier path gets the same rule.
4. No id + several live sessions: refused, candidates named (GIT-103 rule 3). Deliberate behaviour change: a no-id close fails while another agent's session is open on the machine.
5. Payload: sessions/<id>/closing-payload.json wins when present and is refused if it names another session. The shared root closing-payload.json (still what the Stop hook and gitmem init point at) is read for one release when its session_id names the closing session, whatever the number of live sessions; refused if it names another; with no session_id it is used only while a single session is live, otherwise ignored with the "not found at " message. The Stop hook text, gitmem init templates, the agent templates and session_start's payload schema now tell agents to include "session_id", so hook and server agree under concurrent sessions.
6. Result/display show the full id actually written; the raw incoming id is logged to stderr on entry.

Tests (+24 unit in session-close-id-invariant.test.ts; 1 updated; +1 hook test)

Written first and failing on main (13 of the 20 fail there; the others are regression guards, e.g. reclose: true accepted, partial-persist retry, explicit A while bound to B). Covers (incl. root payload with a matching id consumed with two live sessions, mismatched refused, ownerless ignored, and a test that runs the real Stop hook, follows its instruction literally and closes through the server's reader): A + payload carrying B refused; closed session refused / reclose allowed; no id with two live sessions refused; payload mismatch refused; per-session payload read and B's payload untouched; legacy fallback only when one session is live; full id in result and display; stderr log; alias alone, both equal, both different, malformed, logged.

• Unit: 1474/1474. hooks/tests/test-hooks.sh: the new assertion passes; 2 "not detected" tests fail only on a machine that has a gitmem binary on PATH and pass in CI (stock PATH).
• CI: build 18/20/22, hooks-macos, hooks-without-jq green; venue gate (manual dispatch, disposable venue, v1.8.0 schema then current) green.

Known, unrelated

7 e2e tests (user-journey 6, organic-discovery 1) start real claude sessions via the Agent SDK and time out (90s/180s). Not caused by this change; undiagnosed. Full e2e summary lines are in the PR comments.

Not merged, tagged or published; Chris decides. The full rollout of gitmem_session_id to other tools is GIT-103.

🤖 Generated with Claude Code

[nTEG-dev]

Full npm run test:e2e on this branch (GIT-123 sandbox + GIT-122 changes):

 Test Files  2 failed | 8 passed | 2 skipped (12)
      Tests  7 failed | 123 passed | 17 skipped (147)

Same counts as the GIT-123 branch alone (#68). cursor-cross-tool 34/34 passes. The 7 failures are the known, unrelated real-claude timeouts: user-journey (6) and organic-discovery (1). The 2 skipped files are pro-* (no Docker). Sentinel file in the real ~/.gitmem byte-identical after the run.