Set up Cursor Cloud dev environment (uv) + document run notes

[djbhindi]

What this does

Sets up the development environment for this repo in Cursor Cloud and documents how to run the local tooling. This repo (optimizer-seed) is doctrine + memory + thin tooling, not a conventional application: there is no build, no lint, and no automated test suite. The only local executable code is the Python scripts under skills/*/scripts/ (run via uv, with dependencies declared inline via PEP 723) plus the bash memory hooks in tools/.

Changes

• Environment update script (configured via the environment, not a repo file): installs uv to /usr/local/cargo/bin (already on PATH) on VM startup. python3 and git are already present; the scripts resolve their own Python deps (httpx, pyyaml) at run time, so there is nothing else to install.
• AGENTS.md: added a ## Cursor Cloud specific instructions section covering the runtime (uv), which scripts run fully offline vs. which require the external Softmax platform + credentials, and pointing at .runtime/ for scratch work.

Verification (all run in this environment)

• uv installed; all five scripts load and resolve their inline deps:
  • eval_request.py, xp_dashboard.py, fetch_artifacts.py, lifecycle.py, record.py — each --help succeeds.
• Core "record a hypothesis" loop step (record.py, fully offline) exercised end-to-end in a scratch lab under .runtime/:
  • new created 2026-08-27-vote-timing.md
  • validate passed (0 problems)
  • a deliberately invalid record (status: confirmed with empty evals) was correctly rejected with exit code 1
  • list rendered the records table
• Local dashboard web app (xp_dashboard.py) started and served its page + /data JSON endpoint; rendered in Chrome.
• Memory hook tools/rotate_lessons.sh ran as a safe touched-only no-op (no repo mutation).
• Update script re-run confirmed idempotent.

Note: the loop's live steps (hosted evals, league submission) and installing a real game mixin require the external softmax/coworld CLIs, a Softmax account, and network to softmax.com — none of which are provisioned in the cloud VM. Those pieces respond only to --help here; this is documented in AGENTS.md.

Running eval dashboard web app in Chrome

_{To show artifacts inline, enable in settings.}