feat: preserve RLM call lineage across retries

[hallerite]

Summary

• add a generic, harness-bound logical-call resolver for stable identity, private request headers, and namespaced ModelCall.metadata
• replay completed non-streaming calls and coalesce concurrent attempts by opaque logical ID while preserving the existing Stainless retry fallback
• bind every logical ID to route, request bytes, and stable metadata so changed bodies or lineage cannot receive a stale replay
• move every X-RLM-* constant, parser, schema rule, and stripping decision into the RLM harness protocol module
• keep Trace.nodes authoritative for training branches; runtime lineage remains descriptive call metadata

Unrelated harnesses cannot activate RLM behavior merely by sending X-RLM-* headers.

Depends on #2384.

Validation

• uv run pytest tests/v1/test_rlm_lineage.py tests/v1/test_trace.py -q
• complete deterministic v1 suite passed as part of the stack
• Ruff, format, and Ty passed locally and in CI

[macroscopeapp]

🟠 High v1/session.py:102

A cancelled coalesced request cancels the shared LogicalCallReplay.inflight future, so the original exchange cannot publish its result and all other retry handlers receive CancelledError. Await this shared future through asyncio.shield so cancellation of one waiter cannot cancel synchronization for the other attempts.

🚀 Reply "fix it for me" or copy this AI Prompt for your agent:

In file @verifiers/v1/session.py around line 102:

A cancelled coalesced request cancels the shared `LogicalCallReplay.inflight` future, so the original exchange cannot publish its result and all other retry handlers receive `CancelledError`. Await this shared future through `asyncio.shield` so cancellation of one waiter cannot cancel synchronization for the other attempts.

[hallerite]

Superseded by #2386, which now contains the full RLM ACP contract as one PR against main.