Record wave-1 model changes and validation-root correction

[coneilen]

Summary

Update the Windows parity orchestration plan with wave-1 decisions and findings:

• the operator's model change (gpt-6-luna replaces gpt-5.4-mini for small fixes);
• two escalations;
• three environment findings that were causing false local failures.

Changes

• investigation/windows-parity-orchestration-plan.md
  • Models. Small fixes and the start of the escalation ladder use gpt-6-luna, replacing gpt-5.4-mini, by operator decision. The wave-1 evidence is recorded: Windows parity: Right loop panel is readable and matches "THIS LOOP / HANDS OFF TO" #773 was escalated, and Windows parity: Quick Chats canvas and Quick Chat workspace show the chat, not the last loop #774 had a self-caused CI failure. The plan notes that gpt-6-luna has no published agentic-coding benchmark, so it must be calibrated.
  • Wave-1 table. Windows parity: Right loop panel is readable and matches "THIS LOOP / HANDS OFF TO" #773 escalated to gpt-6.1-sol. Windows parity: Quick Chats canvas and Quick Chat workspace show the chat, not the last loop #774 switched to gpt-6-luna. Windows parity: Dialog and sheet framework: in-window dark sheets that size to content and never clip #768 got stuck after 4 capture-harness attempts and is escalated to claude-sonnet-5.5.
  • Validation root. Use a real short directory, not a junction. The alias-versus-resolved path difference breaks the WorktreeGitProcess fixtures, and likely caused the local registered-project failures that CI didn't reproduce.
  • The full unit part needs the desktop lease, because some App and dialog unit roots show windows and take focus.
  • Capture notes:
    • seed the onboarding marker in GRAPHCODE_SUPPORT_DIR;
    • Windows Sandbox isn't suitable for 1440x900 visual captures;
    • create PRs with gh as coneilen.

Test plan

RED: lane A full unit run with -ShellValidationRoot on a C:\gc-a junction -> WorktreeGitProcess cases failed on alias-versus-resolved fixture paths (lane A final-regression log)
GREEN: lane A isolated non-alias WorktreeGitProcess rerun -> 21 cases passed (C:\gcs\a\wgp-rerun.log); Select-String plan -Pattern 'Do **not**' -> 1 match
REGRESSION: git diff --check -> clean; git ls-files --eol -> CRLF preserved; bash Tools/ci/tests/classify-changes.test.sh -> 59 passed, 0 failed

• The RED and GREEN lines come from lane A's logs, not a run in this session.
• Documentation-only change: no product code, so the Windows suites and macOS make test/make check were not run.

Checklist

• I have read the Contributing Guidelines
• I have signed off my commits (git commit -s) per the DCO
• Tests pass locally (make test)
• Code follows the existing style (make check)
• I added the test/contract before the implementation and observed the intended RED failure