fix: expose legacy migration with configured embeddings

[mameikagou]

Summary

• register the documented longmemory migrate CLI command
• pass configured semantic embedding providers into legacy migration instead of silently producing deterministic 8-dimensional vectors
• allow an explicitly empty embedding fallback for fail-closed migrations
• keep post-migration verification read-only so it cannot overwrite persisted sketch dimensions
• migrate legacy project IDs into native LongMemory project hierarchies so MCP project-scoped recall can see imported records
• forward OpenAI-compatible embedding and retry settings through Docker Compose

Verification

• pnpm typecheck
• pnpm build
• pnpm bench:typecheck
• pnpm bench:ci (all gates pass)
• regression migration at 16 dimensions retained 16-dimensional node vectors and sketch state after verification
• production-shaped OpenMemory migration with an OpenAI-compatible provider: 1,263 nodes / 1,640 edges, 0 skipped, 0 errors, integrity benchmark passed
• native project discovery returns all migrated project IDs and project-scoped MCP recall succeeds

Note

pnpm release:check cannot currently start on main because tools/branding.mjs is referenced by package.json but absent from the repository.

[mameikagou]

The container job passes. The validate job currently fails before reaching this PR's changes because package.json invokes tools/branding.mjs, but tools/ is absent from main (same failure reproduces locally). Failed run: https://github.com/CaviraOSS/LongMemory/actions/runs/34318134312

Relevant checks completed locally: typecheck, build, benchmark typecheck, smoke benchmark gates, 16-dimension migration regression, and a 2,560-dimension OpenAI-compatible migration with project-scoped MCP recall.