folder-transfer v0.14.0

Added

• Adaptive compression. Compression is no longer all‑or‑nothing: the server measures, per
  connection, how fast it actually moves data raw vs compressed and only compresses when that is at
  least 25% faster end‑to‑end. On a fast LAN it sends raw (compression there is pure CPU cost
  and roughly halved throughput); on a slow or compressible link it compresses and effective
  throughput rises by about the compression ratio. Fully automatic — -NoCompress still forces it
  off. The wire format is unchanged (the receiver already decodes mixed raw/compressed blocks).
• Benchmark suite (bench/). bench\bench.ps1 builds corpora, drives real transfers through a
  bandwidth/latency‑emulating proxy (bench\bench-proxy.ps1), and writes BENCHMARKS.md; method
  and reproduction steps in bench/README.md. A results table is shown on the project home page.

Changed

• Hot path moved from cmdlets to .NET on both sides (file existence/size/mtime checks, directory
  creation, timestamps, line timing) — a measurable speed‑up on many‑small‑file transfers, in both
  single‑stream and parallel modes.

Fixed

• Destination on an 8.3 short path synced nothing. If -ToFolder resolved to a short path
  (e.g. C:\Users\RUNNER~1\...), the receiver's path‑safety check compared a short root against
  expanded target paths and rejected every file as unsafe. The destination is now normalised the
  same way as the targets.
• Files >= 2 GB in the raw (uncompressed) path crashed with an Int32 overflow in the read loop;
  the byte count is now 64‑bit throughout.
• Small parallel jobs could skip the mirror delete. A stream that lost the connect race and
  claimed no work was treated as a failed stream, marking the whole run incomplete so nothing was
  deleted; such idle streams are now benign and the reconciliation pass runs as intended.