docs: refresh README (accurate positioning + demonstration compiler prominence)

[abrichr]

What

Refreshes the meta-package README so the top-of-file positioning and the Demonstration Compiler section reflect what the ecosystem actually ships now, without adding any capability claims that aren't true.

Changes (README.md only)

• Top positioning: the one-line summary now reflects that OpenAdapt compiles a recorded workflow into a deterministic replay (openadapt-flow) in addition to conditioning/training/evaluating agents, and states the modular optional-extras design as a strength (base install is just the CLI).
• Demonstration Compiler section (added by feat: expose openadapt-flow as the [flow] extra (demonstration compiler under the core brand) #1006) strengthened:
  • notes the standalone PyPI package (pip install openadapt-flow, v0.3.x)
  • sharpens the identity halt into a concrete wrong-record check that refuses a low-confidence match rather than click the wrong target
  • states backend status honestly: reference browser backend shipped (runs in CI, no OS permissions); desktop and RDP backends are adapters in progress, not production paths
  • mentions Skill / MCP emit
  • adds the OpenEMR field result (20/20 compiled vs 10/10 agent, ~half median latency, ~$0.55/run agent vs zero model calls compiled) with its caveats: small sample, shared daily-resetting public demo, not CI-reproducible; CI-reproducible control + adversarial safety measurements published alongside.

Honesty

No new capability claims beyond what openadapt-flow ships and documents. Every metric traces to openadapt-flow's own README/benchmark docs, and every caveat is preserved. Diff is 19 insertions / 4 deletions.

Branch is docs-readme-refresh (the intended docs/readme-refresh collides with the existing remote docs branch — a ref directory/file conflict).

🤖 Generated with Claude Code

https://claude.ai/code/session_01CKrVJJy5jWVCkXAqgUqtqZ