fix(runner): resolve FrameSurface by type so axe-core runs

[tom-sapletta-com]

Problem: testwins/runner.py used hasattr(page, 'frame') to tell a FrameSurface (testwins/frames.py) apart from an ordinary page. Playwright's Page also has frame() (a method), so the condition was always true:

• axe: await page.frame.evaluate(axe_code) failed on a method. Every observation ended with required_detector_failed ("axe-core did not complete"), and every audit with axe enabled was incomplete (exit code 2). This is likely why projects disable axe (for example axe: {enabled: false} in www-heyasa-com).
• url assertions: on an ordinary page they took the FrameSurface branch (page.frame.wait_for_url).

Fix: a _surface(page) helper plus isinstance(page, FrameSurface) in both places.

Tests:

• tests/test_runner_surface.py (written first): a plain page vs a FrameSurface.
• Full suite: 442 passed.
• End to end, an audit of paxlet-com/www-heyasa-com (6 routes × 3 devices, 2 journeys):

	Before	After
axe	0/60 observations	60/60
Scope gaps	60× required_detector_failed	none
Gate	incomplete	passed (exit 0)

• With axe working, real scrollable-region-focusable findings appeared on the site and were fixed there.

Process: the implementation is a willman → koru.develop_ticket patch (PLF-TW-axe, zai/glm-5.3), verified in staging. The repo has no project/new-ticket.sh, so this is an ordinary branch and PR.

🤖 Generated with Claude Code

[tom-sapletta-com]

The red unit job fails in tools/verify_testql.py (task shell-smoke: 1 of 3 steps fails with the testql backend). The same job also fails on main (a20967f and 3c06c7e), because CI installs testql from the moving TESTQL_REF=main. It isn't caused by this PR. The pytest suite (-m 'not browser') passes locally: 442 passed.