refactor: join lineage by request ID

[hallerite]

Summary

This is a focused follow-up stacked on #2449. It keeps that PR's optional ACP lineage manifest and simplifies only the model-call correlation transport:

• store one opaque lineage_request_id on each ModelCall instead of copying session, context, compaction, transition, parent, and depth into a CallLineage envelope
• consume and strip only X-ACP-Lineage-Request-ID at interception; Idempotency-Key remains independent and continues to the provider
• join calls to LineageManifest.requests when ACP publishes its snapshot, with the manifest remaining the sole source of graph structure
• derive calls_by_session, branches_by_session, and branch session/context/compaction IDs through that join without persisting duplicate graph data
• preserve unchanged behavior for harnesses that publish no lineage extension
• pin nano-RLM PR #152 at e26b37a0e8f06d64bff8d7a627ed261be41726a3

The resulting wire contract is:

1. a harness assigns a stable logical request ID;
2. its model request carries the private correlation header (and may independently use the same ID for idempotency);
3. interception records that ID on the call and removes the private header upstream;
4. the ACP manifest maps the ID to its session, context, compaction, and ancestry;
5. verifiers validates and resolves the relationship when it receives the manifest.

Merge order

Merge nano-RLM #152 first, then update this PR's pin to the resulting commit on nano-RLM main before merging. This keeps the exact pin reachable even if #152 is squash-merged and its source branch is deleted.

Validation

• uv run pytest tests/: 919 passed, 76 credential-gated tests skipped
• focused trace/lineage suite: 8 passed
• live Prime RLM ACP resume rollout passed and verifies every recorded call resolves through LineageManifest.requests plus calls_by_session
• two live Prime SWE-bench Verified rollouts passed end to end with forced asyncio.gather recursion and a 1,000-token compaction threshold; each persisted 7 calls = 7 manifest requests, 3 sessions, 4 contexts, one completed compaction, overlapping child-call intervals, complete per-session grouping, and a successful JSON round trip
• exact interception-handler regression covers correlated streaming with and without idempotency, private-header stripping, provider idempotency forwarding, and unchanged ordinary streaming rejection
• ruff check/format, uv run ty check verifiers, touched-file pre-commit, and push hooks: passed

[macroscopeapp]

Approvability

Verdict: Not approved

Macroscope's review found this PR not approvable — This refactor changes ACP request handling, provider-facing idempotency headers, serialized trace records, and removes a public CallLineage API. It also updates the runtime dependency used by existing RLM rollouts, making the impact broader than a mechanical refactor.

Notes:

• No code objects were reviewed. Approvability was decided on eligibility alone.

^{You can add or adjust custom eligibility rules. Learn more.}