feat(verifier): Port/re-implement Quality Verifier functionality from the Python LIP

[matdev83]

feat(verifier): Port/re-implement Quality Verifier functionality from the Python LIP

Reaserch transcript:

Found it. The feature is called the Quality Verifier (the circuit-breaker config still references its earlier codename "Angel" — see the quality_verifier_cooldown_seconds comment in config/config.example.yaml).

What it is

When enabled, every N eligible turns (quality_verifier_frequency, default 10) the proxy holds the main-model completion, calls a secondary verifier LLM (quality_verifier_model) with the conversation history so far, and asks it to audit the last completion. The verifier replies with exactly one of:

• <status>NO_STEERING_NEEDED</status> → original output is returned
• <steering>…</steering> → the proxy does an inline "recall" call of the main model with a private steering system message

It's off by default — enabled by setting quality_verifier_model (CLI --quality-verifier-model backend:model or YAML under session:).

Key functions

Function	Location
QualityVerifierService.is_enabled() / should_run_verification() / should_run_for_request() — scheduling "every N turns"	src/core/services/quality_verifier_service.py
build_verification_messages() — builds the verifier request (prompt + sanitized/truncated session history + completion under audit)	src/core/services/quality_verifier_service.py
validate_quality_verifier_output_format() / maybe_retry_verifier_for_valid_xml() / build_invalid_format_retry_request()	src/core/services/quality_verifier_service.py
Circuit breaker: is_healthy() / report_success() / report_failure()	src/core/services/quality_verifier_service.py
run_quality_verifier_decision() — core decision flow (call verifier → parse verdict → steer/pass/fail-open)	src/core/services/quality_verifier_orchestrator.py
QualityVerifierStreamVerifier.verify_or_passthrough() — streaming path: buffers chunks, verifies, replays or issues inline recall	src/core/services/backend_request_manager/quality_verifier_stream_verifier.py
Turn ledger: _get/_set_quality_verifier_turn_count, reset_quality_verifier_eligible_turn_count()	src/core/services/request_processor_service.py
_extract_quality_verifier_config()	src/core/services/backend_request_manager/streaming_response_handler.py

Most important files

1. src/core/services/quality_verifier_service.py — the service itself (scheduling, message building, format validation, circuit breaker)
2. src/core/services/quality_verifier_orchestrator.py — run_quality_verifier_decision(), the actual verify-and-steer flow
3. src/core/services/backend_request_manager/quality_verifier_stream_verifier.py — streaming integration (verify_or_passthrough)
4. src/core/services/request_processor_service.py — eligible-turn counting/reset (the "every N turns" trigger)
5. src/core/services/quality_verifier_prompt_loader.py + config/prompts/quality_verifier_prompts/ — the verifier role prompt, steering template, tail reminder
6. src/core/services/quality_verifier_steering_messages.py — renders/injects the private steering system message for the recall call
7. src/core/di/registration_helpers/request_processing/_rp_backend_components.py + src/core/services/backend_request_manager_service.py — DI wiring into the request pipeline
8. config/config.example.yaml (session: section) and src/core/cli.py (_warn_if_quality_verifier_frequency_too_low) — configuration
9. docs/user_guide/features/quality-verifier.md — full documentation (also docs/quality-verifier-fractional-turns.md)

The "external LLM providing advice/steering" part is the verifier returning <steering> XML, and the "inline recall" mechanism (append_quality_verifier_steering_system_message) is how the main model actually gets steered on that same turn.