refactor(session): consolidate session and agent runtime loops

[xiami762]

Summary

• Consolidate session execution into SessionLoop, LoopContext, AgentLoop, and StepEngine with explicit preparation, execution, continuation, and finalization boundaries.
• Centralize per-turn runtime prompt context and system-prompt assembly while preserving the existing public prompt-building APIs.
• Align session lifecycle dispatch with the canonical hook stages while keeping legacy hook method aliases compatible.
• Migrate built-in workflow, TUI, and permission configuration references from task to delegate_task, while retaining task as a non-native compatibility alias for older workflows.
• Preserve replay-safe requests, memory bootstrap, model failover, abort handling, queued-input priority, and single terminal-error semantics.

Runtime architecture

SessionLoop
    -> LoopContext
    -> AgentLoop
        -> StepEngine

Prompt assembly

• Build system prompts as ordered SystemPromptBlock values and compose the final provider prompt at the session boundary.
• Carry turn-scoped runtime context separately from stable prompt content so synthetic goal continuations do not reset real-user-turn state.
• Keep Rex and Hephaestus prompt builders compatible while removing the retired task/subtask prompt branches.

Lifecycle hooks and continuation

• Run session.start once and user.prompt.before once per real logical user turn.
• Keep queued real user messages ahead of synthetic goal continuations, including messages that arrive while terminal hooks are running.
• Treat turn.after as an observational terminal hook. Hook output cannot create a synthetic continuation or override a goal that is waiting for user input.
• Keep goal continuation owned by ContinuationPolicy, with abort and step-limit safeguards.
• Fail closed when the LLM before-hook or hook-stage probe fails, so unprocessed requests are never sent to the provider.
• Emit one llm.call.after event per logical model request and aggregate failed same-model retry attempts into its terminal payload.

LLM request and streaming safety

• Freeze the exact provider-bound ModelRequest and reuse it for same-model retries.
• Retain hook-provided stream replacement metadata with the frozen request.
• Restore redacted placeholders in streamed text, reasoning, and reasoning metadata, including placeholders split across chunks and cached-request retries.
• Keep observability inputs redacted while persisting and returning restored assistant output.
• Combine provider-observed prompt usage with post-observation tool-result and later-message deltas before making overflow and cleanup decisions.

Delegation compatibility

• Use delegate_task in the built-in host-forensics workflow, TUI agent flows, prompts, and permission configuration.
• Keep task registered as a non-native compatibility alias that forwards to delegate_task, so older workflows continue to run.
• Canonicalize legacy task permissions to delegate_task in Python and TUI configuration, preserving deny precedence.

Compatibility

• SessionLoop remains the public entry point.
• Toolset and callable-tool selection semantics are unchanged.
• Legacy lifecycle hook method names continue to resolve through compatibility aliases.
• No database schema or migration changes.

Validation

• 92 affected Python regression tests passed after merging the latest dev; four known unrelated existing cases were excluded.
• 21 targeted TUI configuration and provider-usage tests passed.
• Ruff checks passed for all changed Python files.
• git diff --check passed.