test(integration): add checkpoint-backend matrix (I-1)
Commit

The integration suite had no git-refs coverage: every remote-touching test asserted the git-branch v1-branch topology only, while the e2e suite already ran both backends. This adds the missing axis.
TestEnv gains a CheckpointStore field injected as ENTIRE_CHECKPOINTS_PRIMARY into every spawned CLI and git hook (cliEnv + gitHookEnv), and propagated to clones. A ForEachBackend(t, fn) helper runs "git-branch"/"git-refs" subtests, plus backend-aware assertion helpers (CheckpointsPresentLocally/OnRemote, CheckpointExistsOnRemote, RemoteCheckpointState, LatestCheckpointID) mirroring e2e/testutil/backend.go so one test asserts against either topology.
Wires the pre-push and graceful-degradation tests in remote_operations_test.go, the HTTPS push/token tests in http_remote_test.go, and the explain fetch-on-miss test under both backends. Tests whose assertions are inherently v1-branch-shaped (commit counts/subjects, rebase parent counts, checkpoint_remote routing, treeless v1 clone, v1-ref-hash divergence) stay git-branch-only with an explicit rationale pointing at their git-refs follow-up in the test plan.
One genuine git-refs divergence surfaced and is flagged, not papered over: when a session stages its file changes after the stop hook, git-refs condensation creates no per-checkpoint ref (git-branch still creates the local v1 checkpoint). Skipped for git-refs with a KNOWN BUG note; a fresh repo and a stage-before-stop clone both create the ref correctly.
Co-Authored-By: Claude Fable 5 noreply@anthropic.com Claude-Session: https://claude.ai/code/session_012yi3hHGAGepwfrfjPETjGq Entire-Checkpoint: f96568923ebd