You can do this by splitting the task into two tasks - the first task doing the expensive work which is cacheable and the second task adding the volatile data to the output.
In this way the output stays the same and the build cache can be used to avoid doing the expensive work.

                               "' has a reference type ", DataTypeString(dt))
             : absl::OkStatus();
}

// Converts `ninputs` and `inputs` into `inputs_tensors` and `input_nodes` and
// does various checks while doing so. `input_nodes` will contain the same
// information as input_tensors just in a different structure to make
// following processing easier. TODO(iga): Simplify this nested structure.
Status ProcessInputs(

      case CAPSULE:
        return data_.capsule.get() == o.data_.capsule.get();
      case NONE:
        return true;
    }
  }

  /// @brief Implements visitor pattern for doing type-based dispatch.
  ///
  /// @tparam R The desired return type.
  /// @tparam Visitor The visitor class which has a callable operator.
  /// @return The `visitor` called on the correct value.

// overwriting older settings with newer ones.
// If seen != nil, $GODEBUG has changed and we are doing an
// incremental update. To avoid flapping in the case where a value is
// set multiple times (perhaps in the default and the environment,
// or perhaps twice in the environment), we process the string right-to-left
// and only change values not already seen. After doing this for both
// the environment and the default settings, the caller must also call

		// - we want to also handle the case when object is marked as immutable later
		// - Secrets and ConfigMaps are periodically fetched by volumemanager anyway
		// - doing that wouldn't provide visible scalability/performance gain - we
		//   already have it from here
		// - doing that would require significant refactoring to reflector
		// we limit ourselves to just quickly stop the reflector here.
		if c.isImmutable(object) {

    thread.joinSuccessfully(LONG_DELAY_MS);
    assertInterrupted();
  }

  public void testJoinTimeoutMultiInterruptExpired() {
    /*
     * We don't "need" to schedule a thread completion at all here, but by doing
     * so, we come the closest we can to testing that the wait time is
     * appropriately decreased on each progressive join() call.
     */
    TimedThread thread = TimedThread.createWithDelay(LONG_DELAY_MS);

		}

	case *Tuple:
		return subst.tuple(t)

	case *Signature:
		// Preserve the receiver: it is handled during *Interface and *Named type
		// substitution.
		//
		// Naively doing the substitution here can lead to an infinite recursion in
		// the case where the receiver is an interface. For example, consider the
		// following declaration:
		//
		//  type T[A any] struct { f interface{ m() } }

        withParallelThreads(2)

        expect:
        2.times {
            blockingServer.expectConcurrent(":aPing", ":bPing")
            blockingServer.expectConcurrent(":cPing", ":dPing")
            run ":aPing", ":bPing", ":cPing", ":dPing"
        }
    }

    def "tasks are run in parallel if there are tasks without async work running in a different project using --parallel"() {
        given:

	// Signature is valid, lets validate the claims for
	// other fields such as expiry etc.
	return claims.Valid()
}

// ParseUnverifiedStandardClaims - WARNING: Don't use this method unless you know what you're doing
//
// This method parses the token but doesn't validate the signature. It's only
// ever useful in cases where you know the signature is valid (because it has

                if (gradleInstallation == null) {
                    // Store the name without version as well, so we can look it up later
                    // See getNameOfJarReplacedByTestDistributionWithoutVersion why we are doing this
                    Optional<String> replacedPrefix = getNameOfJarReplacedByTestDistributionWithoutVersion(jarFileName);
                    replacedPrefix.ifPresent(prefix -> classpathJars.put(prefix, classpathFile));