fix: scope close-session grouping to one repo when worktree names collide

[SmolSmolStar]

Problem

Closing a terminal with the x close button (or destroy-session) expands to the worktree group so the agent + server terminals die together. getSessionGroupIds queried sessions by both the repo-scoped key (repo-work1) and the bare worktree id (work1).

In a mixed-repo workspace where several repos each have a work1 worktree (the default naming), the bare-key lookup matches every repo's work1 sessions. Clicking x on one terminal destroys every worktree's sessions in the workspace (only differently-named worktrees survive).

Reproduced live on 0.1.22: closing sandbox-work1-claude also destroyed the sessions of agent-workspace/work1, USDUC/work1 and Cat-Code-Academy/work1.

Fix

• getSessionGroupIds: derive repo/worktree via parseWorktreeKey from the session id when the session record lacks them, and query only the repo-scoped key. The bare worktree id is used only when the repository is genuinely unknown (legacy single-repo ids like work1-claude).
• getSessionIdsForWorktree is unchanged, so bare-token lookups from other call sites behave as before.

Tests

• New regression cases in tests/unit/sessionManager.closeSession.test.js: repos sharing a worktree name stay scoped, with and without repositoryName on the session records; legacy pair grouping still works.
• Full unit suite: 103 suites / 603 tests pass.
• Pre-fix code run against the new scenario returns all 8 sessions across 4 repos; patched code returns exactly the clicked worktree's pair.

🤖 Generated with Claude Code

[web3dev1337]

Reviewed as part of an open-PR sweep, side-by-side with sibling PR #1015 (both patch the same lines of getSessionGroupIds). Both were run through a behavioral harness against 4 scenarios: identical results everywhere except one adversarial edge (bare session id + populated repositoryName) where this PR stays correctly scoped and #1015's unconditional bare-prefix second key regresses. This PR also reuses parseWorktreeKey instead of duplicating parsing, and queries one key instead of two. Cherry-picked #1015's richer 4-repo regression test on top so its coverage isn't lost. Verdict: merge this one.