feat(testing): add replay backed provider support for DSL runs

[AdityaShome]

Summary

This PR is rebased on #1566, #1558, #1556, #1555 and #1447 adds replay backed provider support for DSL test runs.

The core change is that mofa test-dsl can now:

• call a live OpenAI and Llama compatible provider
• record the interaction to a tape file
• replay that tape deterministically in later runs

This closes the replay part of the DSL / runner / artifact / replay loop.

Context

Previous work established:

• a DSL entrypoint through mofa test-dsl
• canonical run artifacts
• baseline comparison / comparison output
• CI oriented mismatch handling

What was still missing was a provider path that could:

• record real LLM responses from a live provider
• persist them in a reusable format
• replay them without network access

This PR adds that missing layer.

What Changed

• Added a live OpenAI-compatible provider implementation for DSL-backed recording runs
• Added replay tape types and replay provider plumbing
• Added recording provider plumbing that writes tape interactions to disk
• Integrated recording and replay provider selection into the DSL runner
• Added replay-oriented tests
• Added example DSL cases for hosted recording and tape replay

Files Changed

Main Files

tests/src/live_llm.rs

• Added OpenAiCompatProvider
• Converts kernel requests into OpenAI-compatible chat payloads
• Executes live chat-completions requests
• Converts responses back into kernel response types

tests/src/replay.rs

• Added tape models for recorded interactions
• Added RecordingLLMProvider
• Added ReplayLLMProvider
• Added tape read / write helpers for deterministic replay

tests/src/dsl.rs

• Extended DSL parsing and runner construction to support:
  • live OpenAI-compatible providers
  • tape recording
  • tape replay
• Added provider config loading and API key resolution

tests/src/agent_runner.rs

• Added runner constructors for:
  • live recording providers
  • replay-backed providers
• Wired recording / replay handles into the existing harness

crates/mofa-cli/tests/test_dsl_integration_tests.rs

• Added CLI-level coverage around the replay-enabled DSL execution path

Supporting Files

• crates/mofa-cli/Cargo.toml: Added the testing crate dependency needed for replay-backed DSL execution
• tests/Cargo.toml: Added dependencies required for live-provider requests and replay support
• tests/src/lib.rs: Exported the new replay/live-provider pieces from the testing crate
• tests/tests/dsl_tests.rs: Added DSL tests covering live-provider and replay behavior
• tests/examples/simple_agent_tape.toml: Added a replay-only DSL case that reads a recorded tape
• tests/examples/simple_agent_llama_record.toml: Added a hosted llama recording example through the OpenAI-compatible path
• tests/examples/record_case_gpt_oss_120b.toml: Added a GPT-OSS 120B recording example through the same path
• tests/fixtures/simple_agent_recorded.tape.json: Added a recorded tape fixture used by replay flows

Execution Flow

## Example Commands

Record a live run:

cargo run -p mofa-cli -- test-dsl tests/examples/simple_agent_llama_record.toml --artifact-out /tmp/simple_agent_llama_record_artifact.json --output json

Replay a recorded run:

cargo run -p mofa-cli -- test-dsl tests/examples/simple_agent_tape.toml

Screenshot

Tests

• cargo test -p mofa-testing --test dsl_tests
• cargo test -p mofa-cli --test test_dsl_integration_tests
• Manual recording run against an OpenAI-compatible provider
• Manual replay run using tests/examples/simple_agent_tape.toml

Notes

• The main value of this PR is replay support
• The examples exist to exercise and demonstrate the new recording/replay path
• Provider secrets remain local and are not part of the committed changes