v0.2.1

• Restart watcher, by hand or on a channel start, uses the settings of the last Apply, as
  the README always said. Before, a setting saved but not applied went live at the next
  channel start.
• Observe-only counts toward Switches per hour, so its would_switch lines match what live
  mode would do instead of repeating every 45 seconds.
• The channel and stream lists are read page by page. Before, only the first page was read,
  so past 500 channels or 9000 streams the watcher skipped with "no source after this one".
• The log reader notices a file replaced under the same inode, which Linux allows, instead
  of reading it from the middle.
• The evening of 8 September is a test: its telemetry is a fixture, and the replay must
  find 44 triggers at Stable after 0 and 58 at 180, only on the four starved feeds.
• The README says what Replay counts and when the warm-up starts. Tests, types, lint and
  build run in CI; docs/MEMORY.md holds the decisions, the deployment traps and the
  release routine.

Commits

• 2fe5233 Run plugin actions in the contract tests against a temporary runtime
• 15f093f Read every page of the channel and stream lists
• d0dd8f0 Restart the watcher with the settings of the last Apply
• 2144705 Apply the hourly limit in observe-only mode too
• dc96e2f Say in the README what Replay counts and when warm-up starts
• 462e717 Pin the evening of 8 September in a replay test
• ab61265 Notice a log replaced under the same inode
• a6a6842 Add docs/MEMORY.md with the decisions, traps and release routine
• b52196c Run tests, types, lint and build in CI
• d791cde Release 0.2.1