CI: remaining test failures on costas-code (install scripts, web_server, Playwright E2E)

[SlowGreek]

Context

CI had never run on costas-code — it was gated to push: [main], but main only mirrors upstream and all work lands on costas-code. Three CI runs ever, all on main, none since 2026-07-24.

Fixed during the v2026.7.28 release cycle:

• ci.yml now triggers on costas-code + workflow_dispatch
• contributor attribution (costaspanay@microsoft.com → @SlowGreek)
• apps/desktop/electron/main.ts import-order lint error (hard failure)
• 2 change-detector tests freezing the _load_enabled_toolsets list
• 1 test doing live network I/O (Copilot token exchange blocking the turn thread)
• test_provider_parity (copilot is genuinely dual-auth, like anthropic)
• test_kanban_db + test_update_concurrent_quarantine (asserted the upstream brand string; this fork renders "Catalyst")

Still failing

1. Install-script tests — source-reading antipattern

• tests/test_install_diverged_update.py (2 tests)
• tests/test_install_autostash_conflict_recovery.py (3 tests)

These regex install.sh / install.ps1 source text for a block, and fail with managed-install update block not found.

This is the "Never read source code in tests" antipattern that AGENTS.md bans outright — it fails on a correct refactor and passes on a subtly broken implementation.

Fix properly: extract the update logic into a testable shell function (or a small Python helper) and execute it against a temp git repo, rather than pattern-matching the installer's text.

2. tests/hermes_cli/test_web_server.py

TestWebServerEndpoints::test_status_active_session_count_uses_read_only_db — 1 failure, not yet diagnosed.

3. Desktop E2E / Playwright (Linux)

Failing in CI; not yet diagnosed.

Note

None of these are regressions from recent work — they accumulated while CI was blind to this branch. Release v2026.7.28 shipped with them outstanding, deliberately.