v2.1.0

Added

• examples/land_load.rs and bench/git_push_load.py: landing throughput with N
  concurrent agents, for Pando and git.

Fixed

• bench/concurrent_land.py failed after the rename; README links to the benchmark page
  pointed at a file that still had the old name.
• pandod compact on SQLite no longer leaves a write-ahead log as large as the database
  behind after VACUUM; on git/git that was 392 MB next to a 388 MB database.

Performance

• Concurrent lands to the same branch are batched: the server rebases everything queued as
  one chain and moves the branch with one compare-and-swap. With 200 agents landing at once
  on one machine, throughput went from 10 lands/s with 237 "too contended" errors to 1,012
  lands/s with none; 1,000 agents sustain 504 lands/s (git: 0.2/s). A conflicting change
  fails without affecting the rest of its batch.
• The rebase conflict check walks only the changeset's own paths instead of diffing
  everything that landed since its base.
• SQLite writes are group-committed by one writer thread instead of one transaction per
  write behind a mutex.
• pandod compact is incremental: finished changesets are recorded and later runs process
  only what landed since, instead of rescanning all history.
• Compaction looks up paths in a cache of parsed trees instead of re-reading and re-parsing
  every directory level from storage for each changed file.
• Compaction caches delta chains instead of re-reading a base's whole chain from storage
  for every pair.
• SQLite runs without memory-mapped I/O and with a 64 MB page cache per connection: under
  concurrent writes SQLite remapped the file on every read transaction. On the Linux
  kernel's history, compaction went from 34-47 to 68 changesets/s and import from 78 to 89
  commits/s; clone and log times are unchanged.