Mock-driven parity gate compares GPT-5.4 vs Opus 4.6 without real credentials

Running a model parity comparison used to mean feeding real API keys to a CI job and hoping the budget held. Real providers mean real latency, real rate limits, and real cost — which made the parity gate brittle for regular use. This PR delivers a mock-driven alternative: the GitHub Actions workflow runs both model lanes against a local mock server with staged placeholder credentials, generates the parity report, and uploads the artifacts. No real providers touched.

The mock server gains an Anthropic /v1/messages route alongside the existing OpenAI lane, with proper tool_result ordering in the messages adapter. Mock auth profiles are staged per provider in each agent directory the QA suite uses — the placeholder key qa-mock-not-a-real-key is intentionally not shaped like a real API key to avoid triggering secret scanners.

Ten scenarios now make up the parity pack, with a new instruction-followthrough scenario verifying the agent reads repo instruction files before writing. Second-wave scenarios like subagent-handoff, subagent-fanout-synthesis, and config-restart-capability-flip are guarded by tool-call assertions that require real plannedToolName evidence from the mock's /debug/requests log. Image-understanding scenarios use imageInputCount as their tool-call proxy since vision processing happens inside the provider. Memory-recall stays prose-only by design — the model can pull facts either through a memory-search tool or directly from conversation context, and forcing a plannedToolName assertion would test the harness, not the models.

The parity report now verifies that the run metadata on each summary matches the caller-supplied candidate/baseline labels, throwing a QaParityLabelMismatchError if the paths get swapped. Required scenario failures fail the gate regardless of baseline parity — a scenario that fails on both sides was previously silent. The report title is parametrized so it renders accurately for any provider/model pair, not only the OpenAI versus Anthropic default.

In the qa-lab extension and qa-channel integration points.