[klimovich008]

What this fixes

The answer-domain regression stops compiling on Windows and Linux with use of undeclared identifier 'answers_in'. This blocks source CI for unrelated changes after pipeline-state replay added that dependency to the function under test.

Why merge

• Include the real TLS declaration used by the extracted production function.
• Preserve production branches and every existing assertion; no runtime code changes.
• Keep comparisons with older sources that do not use this declaration working.

Change

• tests/test_program_answer_domains.py: extract the actual answers_in declaration from vk_shader.cpp before the existing function block, after Program is available.

Tested

• python -I -B tests/run_all.py --strict: compiles the fixture and runs its existing default/strict answer-domain cases.
• Fails on the old code, passes now. Both CI platforms reproduced the missing declaration before this fix; both compile and execute the unchanged assertions afterward.
• All eight upstream checks pass: source regressions, Windows build and Linux build.

Not covered

• The fixture exercises normal answer recording; consuming supplied replay answers is outside its existing assertions.

No game data is included.

[klimovich008]

Closing in favor of #1076: it fixes the same compile failure with a simpler fixture declaration and adds coverage for supplied replay answers. Its Windows/Linux source tests and platform builds pass. One PR is enough for this fix.