Rewrite the docs around the managed agent harness concept

[zxgshawn]

The structure was right and the content was not: the pages described what horsie's code does without using the vocabulary this space has settled on, so a reader who knows the field could not place it.

horsie is now introduced as an open-source managed agent harness. Agent = model + harness; horsie is the harness and you bring the model. what-horsie-is maps each primitive onto the component that owns it — harness, agent, environment, session, events, sandbox, checkpoint, trace — and two of those names were already horsie's, which is the point.

The centrepiece is the argument for why this beats a CI job plus an agent process. A CI run is a process: state lives in it, so all that survives is a log, and isolation and execution are the same object, so you cannot pause without destroying the workspace. Here the session, the harness and the sandbox have three lifetimes, so a run can be paused, resumed, and steered — opened, read, answered, and retried one step at a time. That argument now leads the routines page too, since a scheduled run is where it pays off.

Two new pages in How it works: The harness (the loop, why the harness is the only part that decides, inner versus outer harness, guides versus sensors) and Context & memory (context rot, compaction, progressive disclosure, subagents as context isolation rather than parallelism, short- versus long-term memory). Every other page had a vocabulary pass.

what-horsie-is also carries a table of what this category has that horsie does not yet — approval tiers, a secret vault, graded outcomes, versioned agents and memory, OpenTelemetry traces — with what exists instead. Better said by us than discovered by a reader.

No other agent harness is named anywhere, and the prose linter now enforces that. The lowercase codex in the device-login URL and the .claude-plugin/ directory in the bundle format are left alone: those are things you configure, not products being referenced.