docs(readme): launch-funnel pass with evidence table, comparison, and FAQ

[abrichr]

Launch-funnel edit of the Show HN link target, applying the workspace writing guides with minimum effective edits, plus the calibration additions.

Additions

• GitHub-stars badge pointing at the flagship OpenAdaptAI/OpenAdapt repo, labeled as such.
• Quick-links row under the badges: no-install browser demo, website, docs, Discussions, contributing. All five URLs return 200.
• Audited evidence table in the first screen-and-a-half. Numbers copied from the claims registry and the existing Benchmark section: OpenEMR 19/20 effect-verified at 39.2s median (run 20 was a safe halt) vs agent 10/10 at 70.4s median, ~$0.55/run; MockMed 100/100 at 4.9s p50 vs 20/20 at 37.5s p50, ~$0.27/run. Both rows link their method files.
• The existing how-it-works Mermaid now shows independent verification with VERIFIED and HALT terminals, and its text summary matches.
• A condensed agents-vs-RPA comparison table, sourced from the site's /compare data; links to the sourced comparisons.
• A five-question FAQ seeded from the launch reply prep: agents, RPA, failure behavior, secrets/PHI, production readiness (admission-derived wording, no static lifecycle label).
• No Python API snippet: the package exports IR types only, so the honest programmatic surface is the CLI. Skipped rather than faked.

Copy fixes

• Rewrote the four-em-dash evidence-splits paragraph as plain sentences.
• "A workflow program, not just a line of steps" -> "A workflow program."
• "the story isn't success rate. It's that..." -> states the cost/latency difference directly.
• Dropped "honestly" from the substrate-evidence lead-in.

Gates

• python scripts/check_consistency.py passes locally (paths, banned phrases, lifecycle labels, no test count).
• python scripts/validate_claims.py --structure-only passes; no claim text or number changed.
• ci.yml untouched, so the public-artifacts hash pin is unaffected.

🤖 Generated with Claude Code