Fix role-label leakage and add a local Supermemory server

[jhaynie]

Summary

npm run supermemory:up
  read .env                         # gitignored
  if the base URL is Azure Foundry
    start azure-shim                # max_tokens -> max_completion_tokens
    point the server at the shim
  start supermemory-server 0.0.8    # localhost:6767, local embeddings
  write .env.local                  # SUPERMEMORY_API_KEY, SUPERMEMORY_BASE_URL
  add a document and search it

Turns are stored as separate [user] and [assistant] blocks, so the next role label is not part of the memory text. supermemory() passes through the Supermemory SDK client options, including baseURL.

Live evals from upstream pull request #1 are ported onto the memory provider. They were not merged as a git merge because that branch still targets the old extension layout.

Evidence

• Before: a self-hosted boot with no model key exits immediately. With Azure Foundry, memory extraction fails: Unsupported parameter: 'max_tokens'.
  After: npm run supermemory:up against Foundry deployment gpt-5.6-luna printed READY: http://127.0.0.1:6767 stored and searched document yyqns8MAzsPPAEVQaaLgC9. The server log showed memory agent completed (5209ms, 1 memories).
• Before: conversation text user: … best\nassistant: … could be stored as "best assistant".
  After: npm test is 7 pass, 0 fail, including the issue 7 string and client option passthrough.

.env, .env.local, and .env.supermemory match .gitignore (.env*) and are not in this branch.

Merge Danger

Door: two-way

Revert the branch. The Docker volume is local and is not part of the repo.

Blast Radius: provider behavior

New captures use [user] / [assistant] blocks. Client construction now forwards SDK options and reads SUPERMEMORY_BASE_URL. The Docker files are not in the published files list (dist only).