Foundation: simulation and state primitives

Goal

Shared building blocks for recipes that keep state or run a loop, so thirteen advanced recipes do not each invent a queue and a simulator.

A deterministic, seeded simulator base: reset, observe, enumerate legal actions, step, and a trajectory log that can be replayed.
A simulated review queue and action log, used wherever a recipe says an item goes to review or an action is simulated.
A versioned key-value store that keeps provenance and never deletes, with history and rollback.
A transactional wrapper: validate, apply, roll back on failure, with a replayable log.
A budget helper for retry limits, call budgets, and stop conditions, raising when exceeded.
Guidance on using ScriptedBackend to run a closed loop offline.

Each primitive has tests, including property-style tests for the invariants (no deletion, budget never exceeded, replay reproduces state)
Same seed gives the same trajectory on Linux and Windows
A short example shows a loop running offline end to end
Lint and tests pass in CI, or locally with the output pasted in the pull request if CI does not exist yet

Blocked by Foundation: decision backends with typed answers, replay, and scripted modes #63 (decision backends with typed answers, replay, and scripted modes)
Unblocks the recipes labelled needs: simulator
Tracked in Epic: Foundation #74