Make persisted reasoning replay aware of upstream capabilities

Problem statement / motivation

The gateway preserves reasoning items, but its inference preparation assumes vLLM plaintext reasoning. It rejects encrypted-only reasoning and clears opaque state when plaintext exists. That is intentional for vLLM (#151/#153), yet prevents stateful replay to a compatible upstream that can consume its own opaque reasoning state.

Proposed solution

Separate the public stored reasoning item from a capability-aware inference representation. Retain the existing vLLM policy and add opaque replay only for explicitly compatible upstreams. This is a provider-compatibility follow-up, not a request to decrypt reasoning or translate encrypted state across providers.

Acceptance criteria

• Reasoning items, summaries, and opaque fields survive public serialization and persistence without exposing opaque payloads in logs.
• Compatible upstream continuations receive the supported opaque state unchanged and in original item order, including around function calls and outputs.
• vLLM retains its current plaintext normalization and clear encrypted-only rejection.
• Model/provider changes cannot silently replay incompatible state; document routing/identity requirements and return a useful error where compatibility cannot be established.
• Cover HTTP JSON/SSE and WebSocket continuations, stored/manual histories, branching, and compaction boundaries for each supported profile.
• Add regression tests for the existing vLLM behavior and a pinned compatible-provider reference scenario; document the capability matrix under #314.

Additional context

Related: #151/#153 (completed reasoning retention and safe vLLM replay), #314 (capability contract). The vLLM-first deployment should remain unchanged by default. Cross-provider decryption and general encrypted checkpoint interchange are outside scope.

Sources: model capabilities, preserve reasoning across calls.

Reviewed at cdac36f: implementation. Documentation checked September 18, 2026; pin the implementation-time API/model version when recording conformance.

Start with crates/agentic-server-core/src/executor/rehydrate.rs and the capability contract described in #314. Trace how stored and manual histories are prepared for HTTP JSON/SSE and WebSocket continuations, including branching and compaction. Done means compatible upstreams preserve opaque state in order, vLLM behavior remains unchanged, incompatible routing returns a useful error, and the required regression tests and capability documentation are added.