mem migrate silently drops rows with nested idField (e.g. Attio companies id.record_id)

[moekatib]

Summary

`mem migrate` treated `profile.idField` as a flat property name (`row[idField]`) instead of a dotted path. On Attio companies (`idField: "id.record_id"`) where the legacy SQLite stores `id` as a JSON-stringified `{workspace_id, object_id, record_id}`, every row failed to resolve and was silently counted as "skipped" — including on an empty memory store, where skipped is never a legitimate outcome.

Repro

rm -rf ~/.one/mem.pglite
one --agent mem migrate -y
→ {\"platform\":\"attio\",\"model\":\"attioCompanies\",\"rowsSeen\":2024,\"inserted\":0,\"updated\":0,\"skipped\":2024}

Root cause

1. `migrate.ts` used `row[idField]` (flat lookup) instead of `getByDotPath(row, idField)` (same mechanism sync uses).
2. Legacy SQLite JSON-stringifies nested objects before INSERT, so `row.id` was a string like `'{"record_id":"..."}'` — a dotted path wouldn't walk through it anyway without hydration.
3. The report exposed a single `skipped` counter, so a profile misconfiguration and a real key conflict looked identical.

Fix

Shipped in `feat/unified-memory` (next push):

• `reviveStringifiedJson()` rehydrates top-level JSON-stringified columns before id resolution, matching what sync sees live.
• Migrate now calls `getByDotPath(row, idField)` with the rehydrated row.
• Hard reject when the resolved id is a nested object (same guard sync test/mem-writer apply) — prevents silent `[object Object]` collapse.
• Report splits `skipped` into `skippedUnresolvedId` + `skippedError`.
• Human output prints a warning when every row is skipped ("profile missing or idField doesn't resolve").

Verified

Unit tests for rehydration + dotted-path resolution on the exact shape that broke (Attio `id.record_id`).