# test execution

`test/resources.py` owns test concurrency. `test.sh` and `test/quick.sh` source `test/resources.sh` before starting work. That freezes one plan so later suites do not each budget the same remaining memory again. The full gate starts unit and E2E suites together within that plan. Standalone Python E2E and Gleam runners read the same policy.

The planner uses Linux `MemAvailable`, limited by `memory.max - memory.current` for the current cgroup and every ancestor. On macOS it uses free, inactive and speculative VM pages. Unknown hosts receive a conservative 2 GiB budget. A quarter remains outside the test budget. CPU affinity bounds scenarios; Linux CPU quota also bounds active build threads.

The ordinary E2E limits remain four shared Python tests, eight private daemons and four Go scenarios. Half of the shared budget covers E2E scenarios, at 384 MiB per group of one shared case, two private daemons and one Go scenario. The remainder is divided between Gleam modules, Python harness files and build workers. Lower available memory reduces every pool. Less than 1 GiB available fails before work starts. Explicit Python `-j` and `-x` arguments override the plan for experiments. The exported `ALBEDO_TEST_*_JOBS` values can also pin a plan. These are concurrency limits, not hard limits on a test's allocations. The cell-limit scenarios deliberately run 384 MiB child processes.

Gleam runs timed modules first and the remaining modules in bounded EUnit batches. Python harness files run in a bounded subprocess pool. Build limits also set Go package parallelism, `GOMAXPROCS`, Cargo jobs and Gleam's scheduler count. Large machines do not start an unbounded number of compilers or test VMs just because they have many CPUs.

## fixture teardown

A Python test owns every `Albedo` fixture and fake provider it creates. Successful session-creation PUTs record root IDs, independent of later workspace moves. After unittest cleanups, the runner deletes each shared fixture's captured session trees with fresh configuration and family revisions. This stops kernels and background jobs before the shared lane admits another test. A missing root is already cleaned; stale revisions are refreshed up to five times. Other failures fail the test.

A private daemon is discarded synchronously in its worker. Finished tests no longer leave an unbounded queue of daemon-shutdown threads. Provider routes, recorded requests and scripts are released only after session teardown; held provider responses are then cancelled. Successful unittest objects are removed from discovery containers and outcomes retain their names rather than fixtures. Failure diagnostics read a bounded daemon-log tail.

Go session helpers register subtree deletion with `t.Cleanup`. Provider-route cleanup follows session cleanup. Test daemons use explicit idle, grace and preload settings rather than the production ten-minute idle default.