`skills/engineering/tdd` allows for incorrect TDD tests and considers them valid RED -> GREEN

### Skill tdd ### Harness and version Claude Code 2.1.280 ### Model and effort Opus 5 (1M), high effort ### What happened When attempting to replicate an issue, the model decided to fix the issue using a function that did not exist. Failing due to compilation error is only acceptable, if we are not replicating specific issue, otherwise the guard test for the regression never gets to be created. AKA this asserted using TDD what it though the solution should be, not explicitly what the expected functionality (replicate the issue) should be: ``` TDD, verified red → green: 1. Wrote the failing tests first — confirmed red (no method named drain_settled, 4 compile errors). 2. Ported the minimal fix: take_sent/take_acks/drain_settled in counters.rs, and the flush task in lib.rs now calls drain_settled(interval/3) in a re-arming loop. Nothing else. 3. Green: 3 unit tests + 1 integration test pass, and the 3 pre-existing counters_graph tests still pass (7/7). ``` This pass did not replicate the target issue. Tightening of the rules for `bug/issue replication` using TDD could be introduced. ### What you expected A test that would replicate the issue in the existing codebase,... and iterate from there to the fix through the `RED -> GREEN` cycle.