feat: graft recorded sync point so merges ignore squash-stale base

[flipvh]

Problem

Squash syncs advance the recorded sync point (cella.manifest.json + refs/cella/last-sync) but not git's commit graph — a squash commit has a single parent, so git's native merge-base stays at the last true merge ancestor. getEffectiveMergeBase() already accounts for this when computing counts and file classification, but the actual git merge in phase 3 used git's own (stale) merge-base.

The result: every sync replays all upstream changes since the last real merge ancestor. Hunks the fork already integrated — or deliberately resolved away in a previous sync — re-apply as "clean" auto-merges into non-overlapping lines, or re-conflict on files already resolved. Observed in cellajs/raak: a resolved-away scope block in get-attachments.ts reappeared as a duplicated const readFilter on three consecutive syncs, and 8 already-resolved files re-conflicted.

Concrete: raak's recorded sync point was 5 days / 48 commits ahead of git's native merge-base.

Fix

withTemporarySyncBaseGraft() wraps the merge in a local-only git replace --graft that adds the recorded sync point as an extra parent of the fork's HEAD commit. Native git then resolves the 3-way merge against the manifest-tracked base instead of the squash-stale one. The replace ref:

• is deleted in a finally — it never leaks into pushed history (replace refs aren't pushed anyway, but this keeps the working repo clean);
• no-ops when the recorded base is already a native ancestor (fresh merge-commit history, or --hard/--unpinned aggressive mode which intentionally uses the natural base);
• never clobbers an existing replacement object (e.g. the scaffold-root graft from ensureSyncBase).

Wired into both merge paths in merge-engine.ts: the direct-mode phase-3 merge and the analyze-preview worktree merge (replace refs are shared across worktrees).

Validation

• 5 new e2e tests in tests/e2e/sync-base-graft.test.ts reconstruct the raak scenario (fork squash-integrates a sync, resolves an upstream block away). One test is a deliberate control documenting the old behavior (plain merge replays the block); the rest prove the graft prevents re-application, still lands genuinely new upstream work, leaves a real in-progress merge (MERGE_HEAD intact), cleans up on conflict, and respects both no-op guards.
• Verified against the live raak repo: with the graft, git merge cella-upstream/main reports "Already up to date" (previously staged the 18-line duplicate); without it, the block returns.
• pnpm check clean, full suite 137 passing (132 existing + 5 new).

Note

When HEAD is GPG-signed, git prints a "signature will be removed in the replacement commit" warning during the graft — harmless (the replacement is temporary and local), can be suppressed in a follow-up if noise is a concern.

🤖 Generated with Claude Code