fix(graph): migrate inline anchors atomically with grounds_to moves

[abhinav-phi]

Resolves #128. All credit to @tcconnally for the deterministic reproduction and the exact root-cause read — the fix follows their suggested atomic ordering.

Root cause (confirmed against the issue)

persistMovedGroundings processed grounds_to entries first; that branch called deleteGroundedSource(scaffoldFile, oldId) before the inline-anchor pass ran. When a node is grounded both in frontmatter grounds_to and as an inline mex:// anchor, the anchor pass then found:

• getGroundedSource(scaffoldFile, anchor.nodeId) → null (row just deleted), and
• anchorFingerprints.get(oldId) → undefined (snapshot taken from the post-move store),

so if (!baseline) continue; skipped the anchor. The next mex check reported GROUNDING_GONE — permanent until hand-edited, exactly the reporter's step 9.

Fix

Per-node migration is atomic, as the issue suggests: the grounds_to pass records its resolution (oldId → newId) in a per-file map and defers the baseline-row delete; the anchor pass consults that map first, so the anchor rewrites from the same resolution without needing any store state the migration itself just destroyed. The stale baseline rows are deleted only after every rewrite for the file is applied.

One nuance worth review: on current main, a fresh rebuild leaves the old node reachable through node_aliases, which sometimes let the anchor resolve anyway — the bug surfaced deterministically in the reporter's 0.7.1 store where the alias route was also gone. The fix makes the anchor's migration not depend on that coincidence.

Tests

Two new tests in test/graph-integration.test.ts, both following the issue's repro (ground both forms → move with identical body → rebuild → sync):

1. The natural path: both forms migrate (moved = 2), anchor rewritten, no GROUNDING_GONE on the next check.
2. The reporter's store state: old node's grounded-source, fingerprint, and alias rows already gone. Mutation-verified: without the fix the test fails with expected 1 to be 2 (anchor skipped); with the fix both forms migrate.

The 3 pre-existing Windows-env failures in this file (symlink/WAL/atomic-replace) reproduce identically on clean main and are unrelated.

[tcconnally]

Great stuff! Sorry, I kind of gave up on my memory system and the whole Perseus platform in general. Still stoked to see what you continue to build as I think there's definitely room for innovative solutions, just not derivative ones like mine.