[ShotaroKataoka]

Summary

The CI browser job flaked with TimeoutError: DevTools did not start within 15 s on PRs that do not touch the browser check (#625). In all three failed runs it was the job's first Chrome start (test_real_browser_reports_layout_and_js); every later test in the same job started Chrome with a fresh profile in 0.7–1 s, and on passing runs the first test still takes 14–33 s. So the first start on a fresh runner is slow, not hung.

The 15 s was a fixed wait in local_browser, unrelated to the check's own time budget (min(30 + 3 × slides, 110) s, below AgentCore Browser's 120 s session). The failing test had a 48 s budget and gave up after 15 s. The same path runs for a local user's first HTML check after installing.

Changes

• sdpm/sdpm/engine/html/browser.py: a BrowserSource is now called as (width, height, *, start_timeout). The local source passes it to local_browser; a browser that does not start in time is named: Google Chrome did not start within 36 s (no DevTools port).
• sdpm/sdpm/engine/html/browser_check.py: check_deck passes what is left of its budget as start_timeout, so the start shares the budget instead of having its own 15 s. The longest a check can take is unchanged. A check that times out after a start that took at least a third of the budget says The browser took N s of it to start. instead of pointing at a slide script.
• servers/remote/tools/agentcore_browser.py: accepts start_timeout and does not use it (start_browser_session is bounded by the client's API timeout).
• Tests: the start gets the remaining budget, the timeout note names a slow start, local_browser's timeout message and cleanup, the local source passes the timeout; the fake sources take the new argument.
• CHANGELOG.md (Fixed), docs/en/architecture.md: the budget includes starting the browser.

An earlier commit on this branch retried a launch that timed out and added SDPM_BROWSER_START_TIMEOUT (45 s in CI). Both are removed: a retry restarts a slow start from zero, and a longer wait came out of the check's budget, so with 45 s a 2-slide check (36 s) timed out even when the second start worked.

Testing

• With a wrapper that sleeps 20 s before starting Chrome, a 2-slide check is ok (start 21.1 s, 22.0 s in total). Before, it failed at 15 s.
• With a wrapper that never starts, it is failed after 36 s: Browser check could not run (hang.sh): TimeoutError: hang.sh did not start within 36 s (no DevTools port).
• SDPM_TEST_BROWSER=1 pytest tests/test_html_browser_check.py -k real_browser on macOS with Chrome: 13 passed.
• pytest tests/test_html_cdp.py tests/test_html_browser_check.py tests/test_remote_browser.py tests/test_remote_html.py: 92 passed, 13 skipped. Full suite: the only failure, test_html_assemble.py::test_svg_icon_is_inlined_with_prefixed_ids, fails the same way on main in a full run and passes alone.
• make lint

Notes

A start slower than the whole budget still fails, now with a message that names the browser. The CI job is not warmed up on purpose, so it keeps covering the cold first start a user gets.

Closes #625