Fix(eval): sync test-ignore updates from internal reference (#40)

Port the June 2026 test-ignore updates from the internal reference harness
into ProgramBench across 41 instances (+270 ignored tests total), keeping all
metadata verbatim, and update CLAUDE.md accordingly.

Branch-hash keys and per-branch test name-sets are unchanged, so no test
blob re-upload is needed; only ignore metadata differs. The
testorg__calculator fixture is untouched.

Internal-reference-commit: 1ca7719a46b9fc521af2353e287b4f07f4c071ea
Internal-reference-commit: a8f153ad485daf95e2041b5e872b5b715b7f5410
Internal-reference-commit: b6466b9e33c2abd88ce8ca06037c876b1edcc864
Internal-reference-commit: d7f13854873fd23b60606729d649f31a4dc7d7ac
Internal-reference-commit: 7b3a00e057cc625c86bcb9988a439ae50adb0828
Internal-reference-commit: 83488589c5237b4b9af904e1809997df735106f0
Internal-reference-commit: 3bbf4632e7d5e978b817bc168fa0897d0bbc2161
Internal-reference-commit: 691dcaa75a89d8d6f22fc7253c641375df9c3bdf

Author: klieret

CLAUDE.md

@@ -4,6 +4,26 @@
 
 ProgramBench evaluates whether LM-based SWE-agents can reverse-engineer black-box software systems. The workflow: take an open-source CLI tool (mostly Rust/Go), compile it into a Docker image with source removed, then have an LM agent re-implement it from scratch by interacting only with the binary. Behavioral tests (also LM-generated) score the re-implementation.
 
+## Test ignore reasons
+
+Some behavioral tests are unreliable and are excluded from scoring. Each excluded
+test is recorded under `branches.<hash>.ignored_tests[]` in a task's `tests.json`,
+with one or more `reasons[].id` explaining why. All ignored tests are excluded from
+scoring regardless of reason; the id is informational.
+
+- `gold_fail` — test fails **deterministically** on the reference (gold) solution, so it
+  is defective rather than discriminating. Also covers golden-output drift (the gold
+  binary is correct but the captured golden is stale/non-reproducible relative to the
+  build toolchain, an embedded build-stamp, or an external resource).
+- `gold_flaky` — test is **non-deterministic** on the gold solution: it passes in some
+  runs and fails in others. These are timing/race/network/TUI-snapshot flakes, not real
+  defects (distinct from the deterministic `gold_fail`).
+- `dummy_pass` — test passes even on a trivial/dummy executable, so it fails to
+  distinguish a real implementation from a stub.
+- `outcome_dependent_presence` — test appears in some eval runs but not others.
+- `slow_or_hang` — test hangs mid-call or exceeds a duration threshold.
+- `ignored_manual` — manually excluded.
+
 ## Quick reference
 
 ```bash