Add Context Swarm Memory (csm) provider + BEAM 100k result from the omb runner

[muhamadjawdatsalemalakoum]

Context

Following up on the exchange with Chris Latimer about our Context Swarm Memory (CSM) BEAM 100k comparison: the guidance was that official publication requires following the AMB repo instructions and demonstrating results with the benchmark runners. This PR is exactly that, and a question: is an external-provider PR like this your preferred submission path? Happy to restructure however you prefer.

What this PR contains (3 commits, no harness changes)

1. src/memory_bench/memory/csm.py + one registry line — a MemoryProvider backed by the open-source context-swarm-memory TypeScript implementation. initialize() starts a warm localhost bridge service, ingest() streams documents to it (with a durable JSONL log for resume), retrieve() queries it, cleanup() shuts it down. Stdlib-only (urllib/subprocess). Zero changes to AMB scoring, answer prompt, judge prompt, gold data, or any other harness source.
2. README requirements blurb — Node 22+, CSM_REPO_DIR, shared GEMINI_API_KEY, and a Windows note (uv sync --no-install-package uvloop: uvloop arrives transitively via hindsight-api and does not build on Windows; it is unused by this provider).
3. outputs/beam/csm/rag/100k.json.gz — the full 400-query BEAM 100k result produced by omb run on unmodified harness code at base 45fa380, compressed with your own omb publish-results flow (manifest regeneration left to you; we ran from a sparse checkout).

Result (your runner, your judge path)

BEAM 100k	score	correct	avg context tokens	avg retrieve
csm (this PR)	0.743110	337/400	27,026	3.47 s
hindsight (published artifact)	0.733658	326/400	17,654.6	6.38 s

• Answer gemini:gemini-3.1-pro-preview, judge gemini:gemini-2.5-flash-lite — chosen to match the published Hindsight 100k artifact; if you prefer different defaults for chart entries, we will rerun with whatever you specify.
• Single attempt, no resume. Exact command, env, pinned SHAs, sanity captures (omb providers / omb splits), a 1-query smoke, full stdout log, and a 400-row per-query CSM token telemetry sidecar are preserved here: run manifest and result write-up.
• Honest accounting: CSM additionally spends ~8.8k internal input tokens/query inside its own probe/recall/synthesis pipeline (reported separately in the telemetry sidecar so the visible context_tokens never under-states CSM cost), and its answer-visible context is larger than Hindsight's. Retrieval is concurrently faster (warm service + parallel probe/recall stages).
• Integrity: CSM retrieval receives only the ingested documents, the query, user_id, and query_timestamp — no gold answers, rubrics, query IDs, or benchmark-specific logic. We do not describe this as an official leaderboard result unless/until you accept it.

Asks

1. Is this provider-PR + committed-result format the submission path you want, or would you rather run it yourselves / receive something else?
2. If you want different answer/judge models or additional splits (500k/1m/10m), tell us which and we will run them the same way.
3. Small docs note: the README usage examples (uv run amb ... --domain) are ahead of the shipped CLI (omb, --split) at 45fa380 — happy to include a fix here if useful.

[vercel]

@muhamadjawdatsalemalakoum is attempting to deploy a commit to the Vectorize Team on Vercel.

A member of the Team first needs to authorize it.