Fix the Windows parity orchestration process after wave 1

[coneilen]

Summary

Wave 1 of the Windows parity effort merged three product PRs (#788, #789, #792) in about ten hours. Every lane escalated at least once, and the operator stopped the wave. This PR changes how the work is run. The problems were in the process, not in the issues, so the stopped issues (#768, #764, #767, #774, #778) restart from main under these rules.

Changes

investigation/windows-parity-orchestration-plan.md

• Wave 1 outcome: what merged, what was stopped and cleaned up, and a cause-and-fix table.
• Desktop for screenshots only:
  • exact-head CI (three unit shards, integration, packaging) is the regression gate;
  • no local full unit or integration run in desktop slots;
  • one slot of 15 minutes at most per PR, taken last, after code review and green CI.
• Escalation only after 5 failed iterations of the same kind (operator decision), replacing thresholds of 2 or 3. An overclaim, an out-of-lease edit or an idle period is corrected in place, not escalated.
• Validation root of 7 characters or fewer. C:\gc-c2 failed the 81-character preflight by one character.
• Per-PR gate reordered: RED, then GREEN, then a draft PR with exact-head CI, then code review, then screenshots once, then an image review.
• Coordinator model: gpt-6.1-sol, high (operator decision).
• Coordinator rules: verify rulings against the macOS source before sending them; run harness changes end to end locally before merging; answer every lane message within one turn.
• Planning step (operator decision):
  • a planning worker on gpt-6.1-sol, high, investigates each issue and publishes an implementation plan;
  • the coordinator approves the plan and its file scope;
  • then an implementation worker executes it.
• Wave 2 section with this flow.

investigation/windows-parity-lane-kickoff.md (new): one canonical kickoff that every worker reads from main. It replaces hand-copied kickoffs, which drifted: a root that was too long, a wrong Refs line, the wrong reviewer model. It covers:

• the planning and implementation roles;
• ground truth;
• the environment, including every trap wave 1 rediscovered: junctions, root length, the stale profile log, onboarding-seen, PrintWindow versus CopyFromScreen, the ActivateWindow Alt tap (uia-live-gate ActivateWindow Alt tap can leave a modal dialog in system-menu mode #797), zmx cleanup, PR tooling on forks, rebase reverts, and broad gate assertions;
• the gate order, evidence honesty, the 5-iteration rule, and communication.

Test plan

RED: Select-String origin/main plan for "5 failed iterations" -> 0 matches; the plan escalated after 2 or 3 failures and required local full unit runs in desktop slots
GREEN: Select-String edited plan for "5 failed iterations" -> 3 matches; investigation/windows-parity-lane-kickoff.md added with planning and implementation roles
REGRESSION: pwsh Tools/windows/validate.ps1 -Task privacy -> Privacy checks passed; git diff --check -> clean; git ls-files --eol -> plan i/crlf w/crlf preserved, kickoff LF; bash Tools/ci/tests/classify-changes.test.sh -> 59 passed, 0 failed

This is a documentation-only change, 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