v0.2.0

First release since 0.1.130. Headline: a fix for an intermittent end-of-transfer
hang, directory in-band repair, a config file, and compression/CPU performance
wins.

Protocol

• Wire-compatible with 0.1.130 — PROTOCOL_VERSION is unchanged (6), so
  0.1.130 and 0.2.0 peers interoperate over the QUIC/TCP direct path. No forced
  upgrade.

Added

• Config file at ~/.config/mftp/config.toml for persistent per-link
  defaults: streams, chunk_size, no_compress, transport,
  tcp_below_rtt, fec, port. Precedence is explicit CLI flag > config file

  built-in default; an absent file is not an error.

• MFTP_CC environment variable to override the QUIC congestion controller
  (bbr / cubic / reno); the default remains BBR, which suits mftp's
  high-BDP target.

Performance

• Adaptive compression now caps at zstd level 3 instead of escalating to
  level 6 on highly-compressible data. Level 6 cost roughly 2× the CPU for only
  ~7% better ratio, so whenever compression rather than the network was the
  bottleneck it reduced throughput. On a 50 ms-RTT transfer of compressible
  (JSON-log) data, throughput roughly doubled (~67 → ~135 MiB/s) after the
  change. Incompressible data is unaffected (still drops to level 1 / skips).
• Sender skips zero-filling chunk buffers. The file feeders read into
  uninitialised buffers (immediately and fully overwritten by the read) instead
  of allocating vec![0u8; chunk_size] per chunk. Profiling the LAN/TCP path
  showed the redundant zero-fill was ~10% of sender CPU; removed with no
  behaviour change.

Reliability

• In-band incremental repair now covers directory transfers (previously
  single-file only). A missing/corrupt chunk or unreconstructable FEC stripe in
  a directory transfer is repaired over the existing connection — the sender
  reverse-maps the global chunk index across the concatenated file set and the
  receiver scatter-writes it — instead of aborting and resuming. No wire-format
  change.
• Resume files now carry an integrity tag (format v2): a corrupt or
  bit-rotted resume file is detected on load and discarded (clean restart)
  rather than trusted.

Fixed

• Stream-scaling deadlock near end-of-transfer. When adaptive stream scaling
  added streams just as the transfer was finishing, the sender could hang
  indefinitely: if the final worker exited before the scale-up acknowledgement
  arrived, the completion check (which only ran on a worker-join) was never
  re-evaluated, so the loop waited forever on the scale channel and the receiver
  waited forever for a Complete that never came. This hit roughly 1 in 3
  transfers at ~50 ms RTT under the default (adaptive-streams) configuration;
  pinning -n N was a workaround. The dispatch loop now re-checks completion
  after every event, and the case has a regression test.