ci: push fix branches instead of attempting gh pr create

[abueide]

GitHub Actions cannot create pull requests in this org — confirmed by a real run against issue #1309, where gh pr create failed with "GitHub Actions is not permitted to create or approve pull requests." I tried enabling the usual repo-level fix ("Allow GitHub Actions to create and approve pull requests") via the API and got rejected outright: "The enterprise does not allow GitHub Actions to create or approve pull requests." This is a hard block above repo/org admin level.

This applies identically to both claude-doc-sync.yml and claude-issue-investigate.yml — neither had ever actually reached its gh pr create step in a real run before now, so this was a latent bug in both.

Considered routing around it with a dedicated PAT (a human-authenticated token isn't subject to the enterprise's Actions-specific restriction), but decided against circumventing an org-level policy that was presumably set deliberately.

Fix: both workflows now push the fix branch and stop — no gh pr create attempt. Each writes a result file (doc-sync-result.md / investigation-report.md) containing a ready-to-click compare link (.../compare/base...branch?expand=1) so a maintainer can open the PR themselves in one click, uploaded as a job artifact same as before. Also drops pull-requests: write from both workflows' permissions, since neither reads or writes PR data anymore — only contents: write to push the branch.

Follow-up commit, based on that same real #1309 run's output: the investigation prompt now requires a genuine proof of fix rather than an assertion:

• Step 2 asks for a real regression test in the appropriate existing test file (not a throwaway script) that fails on current code for the reported reason.
• Step 4 requires re-running that test after the fix and reporting both results — failing before, passing after — as evidence, mirroring what Claude already did unprompted on the chunk() returns a sparse array when one event exceeds maxKB, stalling the event queue (regression in 2.23.0) #1309 run (it verified its 3 new tests were "not vacuous" by checking they failed without the fix and passed with it).
• Steps 3, 5, 6, and 8 now explicitly forbid presenting a guess as a finding: if the root cause, the fix, or the reproduction is uncertain, the report must say so and state what's needed to resolve it, rather than filling the gap with speculation.

🤖 Generated with Claude Code