[PCrnjak]

Gripper filtering previously shared pose smoothing/cadence, and the simulator applied targets immediately. Add optional independent gripper smoothing/rate settings and an interpolating simulated receiver whose own worker expires commands if the sender stops.

Behavior:

• Omitted gripper settings preserve existing behavior. Explicit time constants/rate caps affect gripper updates without slowing pose output; elapsed-time speed limits and clutch/reference resets remain enforced.
• CartesianReceiver keeps one plan per channel, enforces its own workspace/linear/angular/gripper limits, and uses its monotonic receipt clock for expiry. Duplicate, out-of-order, wrong-session, and invalid packets do not refresh deadlines. Late ticks freeze at the expiry instant; pose-only packets do not extend old gripper/rotation deadlines.
• Explicit arm and fresh clutch re-anchoring handshake with session ID and the sender's last produced sequence. That watermark rejects even unseen packets delayed from before clutching. Expiry requires a new session; construction/import is idle, and close joins the simulated worker even after errors.
• source-leap teleop --demo --interpolate uses the existing packaged browser lab. Omitting the flag keeps immediate feedback. Add callback support, receiver telemetry, a deterministic expiry example, and a focused guide. No robot SDK, physical transport, or additional dependency is introduced.

Validation:

• The installed candidate passed 233 tests on each Windows CPython 3.10.11, 3.11.9, 3.12.10, 3.13.17, 3.14.10, and 3.15.2. Every venv received 30 live sequential and 30 latest-mode frames, replayed the 1,047-frame recording with 50 complete hand samples, ran the interpolated demo, and passed pip check.
• Combined recorded-hand receiver replay produced 18 pose commands and 9 independent gripper updates, respected all receiver motion bounds, completed clutch/resume, and sent zero commands after latched tracking loss.
• A 60-second synthetic soak processed 3,457 control steps and 3,440 simulated targets. Five deliberate sender pauses expired on the receiver worker and required explicit recalibration; reader/receiver threads closed cleanly.
• Wheel-from-sdist build, strict Twine metadata, Ruff, JavaScript syntax, archive-content checks, and 139 internal links passed. A fresh Python 3.10 venv installed the exact final wheel, ran 30 interpolated frames with independent TOML settings, served packaged viewer assets, and passed pip check. Final code/assets match the installed-matrix wheel byte for byte. The runtime pin and base dependencies are unchanged.

Platform CI runs on code commit 693c666192ba9846ccbc09ebc8e913f9670a59d1; subsequent changes are documentation only. See docs/B4_VALIDATION.md and the linked CI checks. Hosted CI verifies software behavior, not Linux/macOS USB/controller support.

Source version is 0.1.0b4. PyPI still provides b3; this PR does not publish a package. No physical robot command, service-state change, or driver installation occurred. UltraGrasp is credited as design inspiration; no external source is bundled by this implementation.