fix(core): retry transient provider compaction failures

[nexxeln]

Summary

• Provider compaction made exactly one llm.compact attempt, so a transient 429/5xx/transport failure on the trigger call failed the compaction and, for automatic compaction, the step. Local compaction already retried these through SessionRunnerRetry.policy and the session.retry plugin hook (fix(core): retry transient compaction failures #47159, feat(plugin): add session retry hook #45999).
• Lift that retry predicate into SessionRunnerRetry.transient(policy, { agent, model, hook }) and use it for both local summary attempts and the provider trigger/endpoint calls. Retrying is safe: nothing is installed until the provider returns a checkpoint, and context overflow is excluded so the existing local-recovery path still handles it.
• Provider compaction now honors the same backoff, retry-after, x-should-retry, and plugin retry hook as every other session request. Docs updated.

Follow-up to #47322–#47324.

Validation

• session-native-compaction.test.ts: a 429 answered once retries under a retry hook (delay: 0, attempt: 2, agent: "compaction") and installs the checkpoint on the second call; persistent failures opt out via x-should-retry: false and still surface provider.rate-limit after one call; cancellation/overflow behavior unchanged.
• Core suite via bun run test: 5,243 passed, 35 skipped; core typecheck clean.