#1446 Responses: replay fixes, experimental persistence and optional summaries

Description

## Summary
Build on main d5ea7133 (0.1.40.3), port stateless replay fixes, and add independently configurable persistence and generated summaries. Both experimental features default off.

## What changed
- responses.py / frontend.py: retain late instruction authority and interrupted reasoning in Responses replay.
- response_store.py: durable response IDs, branching, previous_response_id, retrieve/delete/input_items, retention and restart recovery.
- response_runner.py / server.py: reuse the existing generation queue; add --experimental-responses-persistence and --experimental-responses-summaries. Use --no-experimental-responses-summaries to skip the second generation and retain primary output and raw reasoning.
- Tests and docs cover configuration, API contracts and failure handling.

## Extra Notes
388 regression tests and 128 focused checks passed (overlapping suites; no failures or skips). Includes SDK requests and real process restarts with a mock engine. Persistence stores API history, not KV state. Native GPU latency is not yet measured; this branch is not deployed.