WireDAQ v0.3.0

Version numbers are bumped and reconciled across pyproject.toml, Cargo.toml, and
CMakeLists.txt; tag v0.3.0 to publish to PyPI.

Added

• Per-node clock reconstruction (ADR 0002). The ground station fits a per-node
  ClockModel (online least squares) against its reference clock, recovering each node's
  drift and mapping node timestamps onto one timeline. Recovered drift matches the injected
  drift_ppm within < 0.5 ppm; jitter widens the confidence interval without biasing it.
• wiredaq-clocksync CLI — demos drift recovery and cross-node timeline alignment.
• CsvLogger writes a reconstructed t_ref_us column so drifted nodes align in the export.
• Rust/Lua backend now runs a seeded link simulation — per-packet loss / duplication /
  jitter / reorder with a decode-side receiver, plus executed faults and assertions
  tables; unknown fault kinds are rejected. Its golden vectors are read from the shared
  vectors.json (decode + re-encode) and cargo test runs in CI.
• Hardware-in-the-loop seam. SerialPortTransport (a receive-side ByteStreamTransport
  backed by a real serial device) lets a real board drop in behind the same port the
  simulated serial link used; the existing StreamReceiver → Collector run unchanged.
  Verified without hardware by a loopback test; docs/bring-up-log-template.md captures the
  predicted-vs-observed comparison once a board is attached. pyserial is an optional
  hardware extra — the core stays standard-library-only.

Fixed

• RawFrameLogger and StreamReceiver handle HEARTBEAT frames correctly (were re-encoded
  as / rejected as data, causing byte-mismatched archives and phantom loss).
• Clock drift is no longer quantized to zero at the CLI-default ppm (float accumulation).
• C++ wrapper rejects sample counts that overflow the uint8 wire field instead of
  silently truncating.
• Impairment corruption is always classified as CRC, not framing.
• Decoders (Python / C / Rust) fail closed on control-plane shape: a HEARTBEAT with nonzero
  channel/sample counts is rejected.
• C++ encode() dispatches on msg_type, so encode(decode(heartbeat)) round-trips.
• Collector distinguishes a stale duplicate from a genuine reorder (bounded seen-set), so a
  late duplicate no longer decrements real loss; reorder is counted only on an actual overtake.

Docs

• Fixed stale test counts, softened overclaims ("avionics-grade", "HIL is a drop-in",
  "wired to the schema"), named ADR deciders, and repaired the two broken demo-site links.