Fix verified Windows, CI, and release regressions - #64

Summary

Evidence-driven repair pass over the current preview tree. Every product change has a pre-fix reproduction and a focused regression test; public/private product trees remain identical.

Confirmed product defects fixed

• decode ripgrep event tails as UTF-8 instead of the Windows locale
• make bounded Workspace Git output reads work with Windows pipes
• retry brief Windows sharing violations during atomic image/JSON replacement

CI/release integrity fixes

• fail the Windows multi-command job on the first failed test instead of reporting a false green
• run the stdlib parity checker without importing uninstalled runtime dependencies
• run the wheel smoke outside the checkout so it imports the installed wheel
• refresh checked-in release artifacts

Stale/flaky test contracts corrected

• parse Doctor LOCATIONS as JSON on Windows
• assert platform-aware backend install hints
• make activity ordering deterministic against durable mtimes
• test background work placement rather than runner wall-clock speed
• model operator questions as paused, source handoffs as real Git checkouts, and Reviewer decisions as structured events
• align the Reviewer banner budget with the set of separately enforced contracts
• provide a real byte body in the HTTPError fallback fixture

Verification

• GitHub Actions on this PR: all six checks pass (Linux full, macOS portable, Windows portable, Web/TUI, Windows Desktop, repository parity)
• Local Linux Python 3.11 container: Ruff passed; all 6851 collected tests reached 100% with zero failures
• Web: typecheck + 194 tests passed
• TUI: typecheck + 289 tests passed
• Desktop: Python checks, collection, core tests, TypeScript check, 37 identity tests passed
• PyInstaller frozen backend and provider verification: passed
• Electron unsigned unpacked package: built successfully
• wheel + sdist: built; clean wheel install outside checkout found Web/TUI assets
• final public/private Git tree hash: identical