- **Read-only live regression checks cover the deployed workers.** The health module checks concurrent `/ping` deadlines across configured terminals, bearer authentication, ping log suppression and continued logging on other routes. Each worker now records its actual OpenBLAS thread count at startup, so the live suite verifies native limits rather than only the launcher settings. See [docs/operations.md](docs/operations.md#testing-a-running-stack).

- **API processes no longer commit ~780 MB each for numpy's thread pool.** At `import numpy`, OpenBLAS sets aside a buffer for every CPU of the VM in every API process: 41 GB of commit for 53 terminals on a 40-vCPU VM, before a single backtest ran, and tester agents then ran out of it. `api_runner.bat` now starts the API with `OPENBLAS_NUM_THREADS=1` (and `OMP_NUM_THREADS`, `MKL_NUM_THREADS`), which brought those 53 processes to 3.2 GB. See [docs/operations.md](docs/operations.md#memory-per-api-process).
- **A busy VM is no longer reported hung because of its own healthcheck.** The healthcheck probes every port at once, and each `/ping` wrote two lines through `full.log`'s lock on the Shared folder, so an idle `/ping` took ~1 s on a 45-terminal VM and 5-10 s under load; past the 3 s probe for 10 checks a port counted as hung and the watchdog recreated the VM with its backtests. `/ping` no longer writes request log lines. The log lock is also released only by the write that took it: after a timed-out acquire it belonged to another process, and removing it let a third writer in. See [docs/operations.md](docs/operations.md#busy-is-not-dead--and-hung-is-not-busy).