Implement single and batch agent trace replay

[cursor]

Summary

Replay saved traces against a candidate snapshot/interface to detect behavioral regressions (wrong tool, invalid args, workflow failure).

Milestone

Milestone 10 — Trace Replay

Suggested labels

enhancement, priority:p1, area:cli, area:probes

Acceptance criteria

• tool-semantics replay <trace.json> <candidate-snapshot.json> (names flexible)
• Modes: (1) schema/offline — would prior args still validate / tool still exist; (2) opt-in model re-selection
• Batch mode over a traces directory with aggregate counts (ok / changed / failed)
• Report format distinguishes DETERMINISTIC validation failures vs MODEL-BASED reselection drift
• Exit codes suitable for CI
• Tests with fixture traces + snapshots

Depends on

• Trace schema issue
• Model runner for opt-in mode (Design a provider-neutral model runner #45/Expand provider adapters and packaging for the model runner #60)