fix(coding-agent): adopt upstream OpenCode maintenance session routing

[rynfar]

Prime's OpenCode and OpenCode Go model-provider requests can omit the owning session identity during compaction, refinement, and branch summarization. Ordinary chat can work while those maintenance requests fail.

Review and adopt Prime upstream #2266, merged as 66658d2. It forwards AgentSession identity through maintenance completion calls. This concerns OpenCode as a model provider inside Prime Agent.

Acceptance:

• Carry the correct owning session identity through compaction, turn-prefix summaries, refinement planning/review, and branch summaries.
• Preserve the fork's scoped auxiliary/child request identity and retry semantics; review overlap with feat(agent): stable child-scoped provider identity for RLM subagents, side questions, and auxiliary requests #23 and feat(agent): expose parent session identity to child extension contexts for RLM trees #34.
• Run the local HTTP regression for both OpenCode providers, including stable identity/body across retries, distinct idempotency keys for separate calls, and separation between sessions.
• Verify affected public SDK/Pylon maintenance behavior without weakening transcript continuity or ownership checks.

Suggested priority: reliability fixes first.

Backlog follow-up under #1 and Pylon's Prime integration tracker. The completed September integration remains the baseline; this issue tracks later upstream adoption. Before implementation, refresh upstream evidence and record the ownership decision and focused verification in .pylon/features.yaml and .pylon/upstream-review.md.

[rynfar]

Pre-adoption review confirms the OpenCode routing gap remains: openai-completions.ts constructs the client with options.sessionId before invoking onPayload, so Pylon's scoped payload hooks alone cannot populate x-opencode-session.

Hybridize upstream 66658d2cf73153340f23133fbc16eafac31423e8 with the existing fork hooks. Several upstream positional sessionId additions occupy the same position that Pylon already uses for onPayload; keep both explicitly rather than replacing the fork argument. Preserve distinct summary request IDs, retry-stable request bodies, scoped auxiliary payload identity, and cancellation settlement. Add the owning session routing option to compaction, split-turn summaries, refinement planning/review and branch summaries, then run the upstream local HTTP regressions plus the fork auxiliary-provider/semantic-edge tests.