2.6.0

Added

• Durable branch-scoped RLM history with searchable, paged observations and
  experimental selective context projection. Full session records remain intact.
• One inference ledger and call allowance shared by root, recursive coding
  children and semantic queries, with reported token and USD thresholds,
  restart persistence and explicit unknown-usage reconciliation.
• Experimental Python+Bash RLM profiles for host and Docker. Both tools share
  the persistent environment and sandbox. Command jobs have bounded durable
  output, idempotent admission, deadlines, process-group cancellation and
  explicit recovery when the outcome is unknown.
• RLM settings in :connect, with tool surface, sandbox, observations and
  inference limits. Inspect profiles, family usage, history and command jobs
  while the resident worker is running. A2A remains opt-in and retains its
  separate hosted credit policy.
• A coding profile pilot runner with fresh workspaces, independent graders,
  per-attempt usage and offline execution fixtures. Experimental profiles have
  no demonstrated model quality or cost improvement yet.

Changed

• Native Python remains the default RLM surface. Monty stays a read-only
  research profile; Python+Bash requires host POSIX process groups or Docker.
• RLM run reports include recursive inference usage. Missing provider prices
  and remote A2A usage are shown as unknown instead of zero cost.