feat: add X-Cagent-Compacting header for session-compaction LLM calls

[dgageot]

Session-compaction (summary) LLM calls are currently indistinguishable from ordinary chat completions at the AI gateway layer. This means gateway-side policies — rate limiting, routing, cost attribution — cannot treat them differently from user-facing conversation turns. Adding an explicit marker gives the gateway a reliable signal it can act on, mirroring the already-established X-Cagent-GeneratingTitle pattern.

The change introduces an X-Cagent-Compacting: 1 HTTP header on gateway-bound requests that originate from the compactor. A new WithCompacting() option and Compacting() getter are added to pkg/model/provider/options, and FromModelOptions ensures the flag survives provider cloning. GatewayHTTPOptions in the base gateway is refactored to accept *options.ModelOptions instead of a bare generatingTitle bool, which lets it emit both the title and compacting headers from a single options struct and treat a nil pointer as zero options. The anthropic, openai, and gemini gateway clients are updated accordingly, and the compactor clones its summary model with options.WithCompacting(). The DMR client's skip-reconfiguration guard is extended to cover compaction clones, which also override max_tokens.

No breaking changes. The new header is purely additive; gateways that do not inspect it are unaffected.

[docker-agent]

Assessment: 🟢 APPROVE

The implementation is clean and correct. The X-Cagent-Compacting header follows the established X-Cagent-GeneratingTitle pattern faithfully across all five touch-points: the new WithCompacting()/Compacting() option pair, FromModelOptions propagation, the refactored GatewayHTTPOptions (nil-safe, emits both headers), the DMR skip-reconfiguration guard extension, and the compactor clone site. No bugs found in the changed code.