Fix Codex compaction timeout loop

[rita-aga]

Summary

• use local fallback compaction for openai_codex sessions instead of calling the Codex backend during compaction
• sanitize provider transport timeout text before replying to users
• cap provider-touching Session actions and bound Channel reply delivery timeout

Verification

• cargo test -p temperpaw --quiet
• cargo test --manifest-path os-apps/paw-agent/wasm/context_compactor/Cargo.toml --quiet
• cargo test --manifest-path os-apps/paw-agent/wasm/agent_reply/Cargo.toml --quiet
• cargo test -p temperpaw --test session_turn_architecture --quiet
• production Railway deployment verified healthy; Datadog post-deploy failure signatures count is zero