Fix: affected-test selection missed tests on multi-targeted projects

[csa7mdm]

Bug

dotnet_test_affected could miss tests that call the changed code directly while still reporting selectionComplete: true. The result also varied between sessions.

Found by: the Wave 0 benchmark gate. On Polly commit 8ba491d7, a change to RandomUtil.cs:

• selected 121 tests in one session and 118 in another;
• the 118-test runs missed RandomUtilTests.Next_Ok and NextDouble_Ok;
• fresh sessions on main gave 118 every time.

Cause:

1. A multi-targeted project loads once per target framework, so a changed file exists as several documents.
2. The walk seeded from the first in-scope document id and deduplicated visited members by documentation id.
3. The search from one framework copy's symbol doesn't reliably reach the dependents of the other copies. Roslyn links copies by source position, and #if branches put the same member on different lines.
4. Which copy came first wasn't stable, hence the session-to-session flip.

Fix

• Seed from every in-scope copy of each changed file.
• Track visited members per (project copy, documentation id). Tests are still reported once.

Evidence

Unit test: a new test reproduces Polly's shape: an internal class with #if NET / #else, one test reaching each framework copy.

• On main it fails in both orders (each finds only its own copy's test).
• With the fix it passes both.

Replay: Polly's last 40 source commits, sequential, maxSelectionSeconds: 120, compared by test name on the 22 commits complete in all runs.

Build	Lost vs main	Gained vs main	Median / p90 selection	Filtered runs
main	-	-	0.81 s / 2.24 s	17
this PR	0	14 (commit 708b02b0)	2.82 s / 6.05 s	16
rejected variant: per-copy keys only for #if files	32, 62 and 56 tests on 3 commits	14	1.17 s / 2.77 s	16

The rejected variant is why this PR keys every member per copy. The faster version wasn't safe. RandomUtil.cs now selects 121 tests on every run (6 of 6 in a fresh session).

Cost: selection is about 3× slower on multi-targeted solutions. A few more broad changes will fall back to whole test projects within the default 10 s budget. Correctness first; the persistent index planned for Wave 2 addresses speed.

Tests

166 in the solution pass (Testing 64, including the 2 new cases).

🤖 Generated with Claude Code