Improve session backup observability and latency

[guima-why]

Summary

• log session startup configuration and backup duration with session IDs
• reduce pipeline backup latency by removing per-step-completed backups, simplifying backup mirroring, and preserving critical recovery ordering
• select real E2E models by input type: deepseek-v4-flash-0731 for text scenarios and qwen3.8-max for multimodal scenarios

Why

Production backup storage may be mounted on OSS and can add substantial latency before the step 4 candidate selection UI appears. These changes make backup costs observable, remove redundant work, and retain the critical backups needed for recovery across sequential sandbox moves.

Behavior

• A2A input_required/waiting_input is published before backup; terminal and handoff states remain gated by critical backup
• non-A2A pipeline input_required retains its critical backup
• normal REPL, Web, ACP, headless, process-mode, and A2A session starts log effective runtime parameters without exposing endpoint paths or credentials
• explicit E2E --model still overrides automatic per-scenario routing

Validation

• make test: 13,481 passed, 2 skipped
• make lint: Ruff and ty passed
• real A2A E2E: scenario1, scenario1-performance-backup, ask-waiting passed
• real REPL E2E: scenario1, selection-waiting-resume passed
• subagent and Claude Code compatibility reviews found no incompatibilities with current main
• git range-diff confirmed all three feature patches were unchanged by the rebase onto origin/main