Synced from cloud PR: Skyvern-AI/skyvern-cloud#8990
Author: @pedrohsdb

## Summary
- Integrates adaptive caching into the workflow execution lifecycle and script page class — split from #8960
- This is the main production-touching PR — connects all the standalone modules together
- Depends on #8984 (database), #8986 (prompts), #8987 (modules), #8988 (SkyvernPage), #8989 (script gen)

## Changes (797 additions, 78 deletions)

### `skyvern/forge/sdk/workflow/service.py` (+765 lines)
The core orchestration changes:
- Handles `code_v2` run mode in the workflow execution path
- Triggers script review on workflow completion when adaptive caching is enabled
- Manages fallback episode collection throughout the run lifecycle
- Coordinates with the script reviewer service to publish improved script revisions
- Adds `generate_script_on_terminal` support for auto-generating scripts on terminal workflow states

### `skyvern/core/script_generations/script_skyvern_page.py` (+110 lines)
Extends the script execution page class:
- Records branch hits during `classify()` calls for TTL-based pruning
- Captures fallback episodes when cached script logic fails
- Routes to `surgical_fallback()` when targeted AI intervention is needed

## Test plan
- [x] Both files pass `py_compile`
- [x] Both files pass ruff lint and format
- [ ] End-to-end test: run a workflow with adaptive_caching=true
- [ ] Verify fallback episodes trigger script review on completion
- [ ] Verify branch hits are recorded and stale branches detected
- [ ] Test generate_script_on_terminal flag behavior

🤖 Generated with [Claude Code](https://claude.ai/code)