test: bound all-crate spawners and add a hang backstop to nextest

.config/nextest.toml

@@ -10,9 +10,19 @@
 # Local runs: no retries, so a flaky failure surfaces instead of being masked.
 retries = 0
 
+# Determinism backstop (WOR-1405). The deadlock this file guards against turns a
+# starved child-spawning test into an indefinite hang, which freezes the whole
+# local run with no output and no end. The concurrency bound below removes the
+# starvation, but a backstop guarantees the run always terminates: any test
+# still running after 180s (three 60s periods) is killed and reported failed
+# rather than hanging forever. No default-feature test comes close to 60s, so a
+# healthy run never trips this; it only fires on a genuine hang or regression.
+# Profiles inherit unset settings from `default`, so ci gets the same guard.
+slow-timeout = { period = "60s", terminate-after = 3 }
+
 # Bounded concurrency for the process-spawning integration tests (WOR-1405).
-# 80 of the 87 `mcptest` integration-test files fork a `/bin/sh` JSON-RPC
-# fixture or spawn an `mcptest` child and assert on a round-trip. Each such test
+# Most integration tests across the workspace fork a `/bin/sh` JSON-RPC fixture
+# or spawn an `mcptest`/server child and assert on a round-trip. Each such test
 # needs CPU for both itself and its child, so at nextest's default parallelism
 # (one test per core, thousands queued) the cores get oversubscribed two-to-one,
 # the children starve, and a blocking round-trip read hangs the whole local run
@@ -46,13 +56,18 @@ path = "junit.xml"
 filter = 'test(repl_prints_notifications_and_fulfills_sampling) | test(graceful_shutdown_terminates_child)'
 threads-required = 'num-cpus'
 
-# Assign every `mcptest` integration test (the `tests/` dir, kind = test) to the
+# Assign every workspace integration test (the `tests/` dir, kind = test) to the
 # bounded `stdio-spawners` group so the swarm of child-spawning tests cannot
-# co-saturate the CPU (WOR-1405). Nearly all of them spawn a child; scoping by
-# `kind(test)` is both robust (no hand-maintained list to drift) and correct
-# (unit tests stay fully parallel). The two tests above also match here, which
-# composes cleanly: they reserve the whole pool when they run, which already
-# satisfies this group's bound.
+# co-saturate the CPU (WOR-1405). The child-spawners are not confined to one
+# crate: the bulk live in `mcptest`, but `mcptest-core`
+# (stdio_transport_integration, auto_discover_doctor_corpus) and `mcptest-agent`
+# (the agent end-to-end suites) fork children too, and a `package(mcptest)`
+# filter left those unbounded. Scoping by `kind(test)` alone is both robust (no
+# per-crate or per-test list to drift as suites are added) and correct (unit
+# tests, kind lib/bin, stay fully parallel). The genuinely pure-logic
+# integration tests it also catches are fast, so bounding them costs little. The
+# two tests above also match here, which composes cleanly: they reserve the
+# whole pool when they run, which already satisfies this group's bound.
 [[profile.default.overrides]]
-filter = 'package(mcptest) & kind(test)'
+filter = 'kind(test)'
 test-group = 'stdio-spawners'