Fix recording targets and harden quickstart evidence

[abrichr]

Summary

• apply macOS and local RDP/Citrix recording target flags to the exact Capture window
• refuse Windows agent, Linux target, network RDP host, cross-surface, and conflicting record flags before Capture starts when the current recording schema cannot bind them
• remove the short-lived Playwright driver probe that emitted unhandled TargetClosedError warnings after successful commands on Linux, macOS, and Windows
• make clean-machine evidence fail on unhandled asynchronous runtime errors and bind evidence to the exact source revision and wheel hash
• correct the claim registry so browser end-to-end evidence names the required e2e-browser PR gate and the weekly compatibility matrix
• replace the stale ecosystem memo and expand the README and surface guide with the current Capture Beta role and exact install/record/compile/replay commands

Root cause

The shared record parser accepted replay-only target flags, but the desktop record path did not apply them. A command could therefore appear to record one target while Capture observed another. The browser readiness probe started Playwright only to inspect an executable path; Playwright 1.62 could leave its driver task pending and print a false-success-shaped exception after an otherwise healthy run. Generated claim documentation also retained the retired nightly-gate wording.

Validation

• 173 focused CLI, recording, browser setup, lifecycle, workflow, and claim-registry tests passed
• Ruff lint and format passed
• Flow wheel and source archive passed the release and license boundary
• clean installed Flow lifecycle passed record, compile, expected strict-lint refusal, permissive certification, expected clinical-policy refusal, baseline replay, drift replay, Standard tutorial verification, and uninstall
• combined installed launcher + Flow quickstart passed: VERIFIED, 2/2 independent effects, 0 model calls