fix(eval): match JUnit names ignoring optional eval. prefix

[rootkiller6788]

Problem

The evaluator compares the expected test list in tests.json against JUnit classnames literally, but the two disagree on the package root. tests.json is inconsistent: some tasks store eval.tests.foo (csview) while others store tests.foo (cmatrix, the bundled calculator fixture). JUnit classnames are always rooted at eval.tests.foo because pytest collects the suite from /workspace/eval.

For tasks whose metadata omits the prefix, _process_branch_xml therefore treats every test as both:

1. missing from the XML, injecting a synthetic not_run record, and
2. unexpected in the XML, emitting a "not in tests.json" warning.

A branch that fully passed is counted twice and scores half. The bundled fixture data/test_runs/correct/testorg__calculator.abc1234 reproduces this exactly: three passing eval.tests.* records plus three not_run tests.* records. The same literal comparison in get_branches_to_eval makes a completed branch look incomplete, so programbench eval re-runs it on every non---force pass.

Fix

Compare test names in a canonical form that strips an optional leading eval. prefix from both sides. Added _canonical_test_name and applied it to the expected/observed matching in _process_branch_xml and get_branches_to_eval. Parsed result names are left unchanged, so existing eval.json files and ignored-test filtering are unaffected.

Verification

• New regression tests in tests/test_eval.py (namespace mismatch matches, missing test still not_run, re-eval idempotent) fail on main and pass with the fix.
• Full suite: 68 passed.
• ruff check and ruff format --check clean at 0.15.7.

Refs #56