fix(game): use the original Borland random stream

Replace the Xorshift state with the reconstructed Borland Win16 generator,
including its wrapping linear update and high-word Random(n) mapping. Route
launcher clamp variation, effect selection, and claw terminal choices through
one shared stream so seeded event order follows the executable.

Use the same stream for the recovered magnetic-field response. Fields now scale
horizontal speed by 0.9 and choose the upward impulse from
-3800 * (1 - Random * 0.3) instead of assigning a fixed -3000 velocity. Adjust
deterministic scenario expectations where the original stream permits a second
natural claw capture or does not force one during a particular autoplay seed.

Test Plan:
- `cargo test --all-targets` -- passed, 56 tests
- `cargo clippy --all-targets -- -D warnings` -- passed
- `rumdl check CHANGELOG.md RECONSTRUCTION.md` -- passed
- `git diff --cached --check` -- passed

tdkpin-rs/src/simulation.rs

@@ -389,8 +389,6 @@ mod tests {
         assert!(event_count("Bumper") >= 1);
         assert!(event_count("Target") >= 1);
         assert!(event_count("Lock") >= 1);
-        assert!(event_count("ClawCapture") >= 1);
-        assert_eq!(event_count("ClawCapture"), event_count("ClawRelease"));
         assert!(event_count("Drain") >= 1);
         assert!(simulation.trace.iter().all(|snapshot| {
             snapshot.ball.x.is_finite()
@@ -459,7 +457,6 @@ mod tests {
                 "{} must release every captured ball",
                 scenario.name()
             );
-            assert!(!simulation.game.claw.active);
             assert!(!simulation.game.claw.ball_suspended);
             assert!(simulation.game.ball.position.is_finite());
             assert!(simulation.game.ball.velocity.is_finite());