[Hermes][Browser tests][1/n] Prevent real Chrome launch in retry test

[OmarB97]

Why

test_browser_manage_connect_default_local_retries_after_launch still mocked the retired boolean try_launch_chrome_debug helper after the gateway switched to the structured launch_chrome_debug API. Running the test therefore launched a real detached Google Chrome with a pytest-temporary profile; repeated test runs opened windows, stole focus, and surfaced the fixture text partial answer complete on the live desktop.

What changed

• Mock hermes_cli.browser_connect.launch_chrome_debug, the symbol the gateway actually imports.
• Return ChromeDebugLaunch(launched=True) so the test exercises the production structured result contract without starting an OS process.
• Production browser behavior is unchanged; this is a regression-test isolation repair.

How to review

Read the single hunk in tests/test_tui_gateway_server.py and compare it with tui_gateway/server.py where _browser_connect imports and calls launch_chrome_debug. Confirm the replacement mock matches that call and preserves the existing retry assertions.

Evidence

Before the fix, the machine accumulated eight detached Chrome roots with command lines containing pytest-of-obaradei/.../test_browser_manage_connect_de2/hermes_test/chrome-debug; another unfixed run created PID 14501 at 2026-07-12 13:40:05 PDT and a later run created another orphan five minutes later. After the fix, both the exact test and the adjacent browser-connect slice completed with test_chrome_before=none and test_chrome_after=none.

Verification

• scripts/run_tests.sh tests/test_tui_gateway_server.py -k test_browser_manage_connect_default_local_retries_after_launch -> 1 passed, 0 failed.
• scripts/run_tests.sh tests/test_tui_gateway_server.py -k browser_manage_connect -> 16 passed, 0 failed.
• Process-set assertion around both runs -> no Google Chrome process with a pytest test_browser_manage_connect_de profile was created.
• git diff --check -> passed.

Risks / gaps

• Accepted risk, no follow-up: only a stale test mock changes; production launch code is untouched.
• Already-running agents on old checkouts can still launch one final orphan until their current test loop is stopped or updated. Immediate local cleanup and live-loop tracing are tracked in MeshBoard task hermes-browser-connect-test-no-real-chrome-20260712.
• No follow-up needed: upstream delivery uses a separate clean branch based on NousResearch/hermes-agent:main because the fork and upstream currently have no merge base; this fork PR intentionally contains only the fork-main version of the same one-hunk repair.

Collaborators

Participants:

• @OmarB97 — operator on ko-mac
• GPT-5 Codex — OpenAI, interactive implementation lane

Process:

• Time to finish: about 25 minutes
• Iteration: 1 of 1
• Lead reviewer: MeshBoard peer-review lane (exact-head review pending)

Task context:

• Task: hermes-browser-connect-test-no-real-chrome-20260712
• Size: S · Risk: medium · Priority: High

Related work:

• Diff review: hermes-browser-connect-test-no-real-chrome-20260712-review
• Upstream PR: opened from clean upstream branch after this fork mirror