Make script/test fail on flaky failures and enable -race (closes #32)

script/test ended with `go test ./... || go test -v ./...`, so the
verbose rerun's exit status became the script's. A test that failed
once and passed on the retry produced exit 0, and since the Dockerfile
runs `make test` and CI runs script/cibuild, flaky failures were
invisible repo-wide.

Replace that with the pattern from REPO_POLICIES.md: run
`go test -timeout 30s -race -cover ./...`, and on failure rerun
verbosely and then exit 1, so the rerun is diagnostic only and can
never turn a failed run green. `go vet ./...` still runs first; the
script stays POSIX sh with set -eu, the CGO_ENABLED=1 export, and the
repo-root cd idiom.

Enabling -race surfaced a genuine timing conflict rather than a data
race: no package that ran to completion reported one, but internal/cli
now exceeds the 30s timeout because the 99MB/100MB boundary cases in
secrets_size_test.go are about 10x slower under race instrumentation.
Measured in the CI-equivalent container (script/cibuild, with the
memlock ulimit): 18.7s for the make test layer without -race, versus
internal/cli alone needing roughly 55-60s with it. Neither the flags
nor the tests were weakened to hide this; the conflict is filed as #52
for a decision.

script/test

@@ -9,7 +9,11 @@ main() {
     # CGO is required (Makefile exports this too)
     export CGO_ENABLED=1
     go vet ./...
-    go test ./... || go test -v ./...
+    # The rerun is diagnostic only: `exit 1` keeps the script failing
+    # even if a flaky test passes on the second attempt.
+    go test -timeout 30s -race -cover ./... || \
+        { echo "--- Rerunning with -v for details ---"; \
+          go test -timeout 30s -race -v ./...; exit 1; }
 }
 
 main "$@"