perf(embeddings): speed up and harden index builds

[srobinson]

Summary

• Reuse unchanged semantic vectors during structural force refreshes and add --force-embed for explicit full semantic rebuilds.
• Run embedding requests with configurable bounded concurrency, retries, delays, and timeouts.
• Split sections above the OpenAI input limit into token safe inputs, preserve document and heading context, then pool them into one normalized section vector.
• Keep each request below the aggregate token limit and report the affected document and heading when chunking occurs.
• Update the API enabled semantic integration fixtures to use the current generation session contract.
• Include the requested agent coordination additions in LESSONS.md.

Test plan

• just test: 99 files passed, 1,778 tests passed, 11 skipped
• pnpm typecheck
• pnpm build
• 50 focused CLI, config, batching, token, persistence, and manifest tests
• Oversized section integration test confirms every API input stays within 8,000 tokens and one vector is stored
• Production run embedded 6,072 sections in 29.7 seconds with no input length failure
• Installed CLI exposes mdm index --force-embed