cendor-sdk 1.15.0

Phase-S follow-up — the parity items deferred from 1.14 land, both languages, on the same
cendor-core 1.10 / cendor-tokenguard 1.5 shelf (no new floor).

Added

• Streamed + team conversation.id (S6): run.stream/run.astream and run_agents/run_agents_async now stamp Result.conversation_id from a keyed session and enter the conversation scope, so a monitor groups multi-turn streamed and multi-agent runs — previously single-agent-blocking only.
• Streamed checkpoints (S13): run.stream/run.astream accept checkpoint= (a path or Checkpointer) — per-turn saves for a single agent, per-turn + per-segment for a team. A finished checkpoint replays a lone terminal RunComplete (no model call, no re-yielded deltas); an unfinished resume continues from the saved messages without re-emitting prior deltas.
• Bedrock forced-toolChoice structured output (S14): a tool-less Bedrock agent with output_type now forces a synthetic schema-shaped tool (toolConfig.toolChoice), and the provider unwraps its input into the final answer — stronger than the JSON nudge. Gated to tool-less agents (a forced choice can't coexist with real tools on Converse); with tools present it falls back to the nudge (documented).

Changed

• Streamed-run RAG attributed (S11): the four stream generators now prepare messages inside the run scope (collector + trace + per-agent budget), so an always-on retriever's embed call is attributed to the run rather than firing before it opened — parity with the blocking run() (GLR-4). No API change.