v0.11.0 adds chat.BoundConversation, the seam a provider applies a history
policy through, and the conformance case that proves it runs.

The case fails on this bump alone — history_policy_applied reports the last
request carrying 9 turns with TruncateOldest(2) set. The policy has never been
applied here, and until now nothing said so.

Config.HistoryPolicy has had no effect on this provider since it shipped. A
caller setting TruncateOldest(40) got a conversation that grew without bound.

Before every request the client now describes its retained turns, passes them
through chat.BoundConversation, and rewrites the session from the indices
returned. Before every request rather than once per call: the ReAct loop
appends a tool request and its results per step, so an entry-time bound never
sees the transcript that overflows. All three paths run it — Chat, Ask and the
streaming step.

Two things are specific to this provider:

  - The system turn lives in the message list here, not in a separate field.
    It is configuration rather than conversation, so it is held out of the
    described set entirely instead of being described and defended. A policy
    cannot drop what it was never offered.
  - OpenAI has no per-message cache annotation. It caches eligible prefixes
    automatically, so there is nothing to read back off a turn: the cached
    region is a length. AddCached now records it, and those turns are described
    as Pinned. Rewriting any of that prefix costs every later cache hit.

The trailing tool unit is one assistant turn plus every tool turn answering it,
because OpenAI sends one tool message per call made. Holding out only the last
of them would leave the rest in the described set, where a policy is free to
drop them and split the group.

Refs chat#13