fix(agent-core-v2): resolve reopened agents to their roster branch - #3726

Related Issue

Internal bug report (no public issue): reopening an agent after undo + process restart resurrects the undone turns.

Problem

SessionStores manages one branch per agent plus a _session branch whose agent.opened / agent.switched / agent.closed events fold into a roster mapping each agent to its current branch. This roster must exist because undo is non-destructive: undo('main', 1) forks a new branch main~2 and switches to it, while the old main branch keeps its full pre-undo history.

The read side had a gap: open() resolved branches purely by agent name (tree.has(agentId) → open the same-named branch) and never consulted the roster. Within the same process this went unnoticed (the in-memory store was already reset), but after a process restart open('main') opened the stale pre-undo branch and every undone turn came back. The migrate-v2 end-to-end test asserted only the roster contents after a restart, never the history the reopened agent actually sees — exactly the gap.

What changed

• SessionStores.open() now treats the roster as authoritative: it reads the folded roster first and opens the registered branch when an entry exists (a roster entry pointing at a missing branch fails loudly via openBranch — corrupted data, no silent fallback). Without an entry it keeps the existing fallback — open the same-named branch (covers the crash window between createBranch and the persisted agent.opened) or create one from opts.from — and dispatches agent.opened to backfill the registration whenever it opens a branch with no roster entry, so first open and crash-window repair share one registration path.
• Tests extended with zero net test-count change: the stores reopen case now undoes before reopening and asserts the restored agent follows the roster (main~2, post-undo history) while the fork branch keeps its own full history; the migrate-v2 end-to-end case now also opens the agent after the simulated restart and asserts ref.branch and the folded history, not just the roster contents.
• No changeset: SessionStores is only consumed by human/ tests and persist/open.ts so far — production DI wiring lands in a follow-up task, so the fix is not user-perceivable yet.

Checklist

• I have read the CONTRIBUTING document.
• I have linked a related issue (external PRs: the issue must have a maintainer's /approve).
• I have added tests that prove my feature works.
• Ran gen-changesets skill, or this PR needs no changeset.
• Ran gen-docs skill, or this PR needs no doc update.