feat: Qualification dual-agent loop — auto-correct agent from test failures

[VictorGjn]

Problem

The Qualification tab exists (545 lines backend) but has never been verified e2e with the v2 wizard. The dual-agent loop (Agent Testeur + Agent Correcteur, Samuel Neveu pattern) from ROADMAP B1 is the strongest unshipped differentiator.

Success criteria

• Qualification tab works e2e: generate suite → run → results → patches
• After run with score < passThreshold: "Agent Correcteur" analyzes failures and proposes system prompt patches
• Patches applicable in one click. Max 3 correction iterations.

Non-goals

No infinite loop. No unsupervised auto-apply.

Implementation

1. Run existing integration tests (qualification-api.test.ts, qualification-e2e.test.ts) and fix breakage
2. Wire QualificationTab.tsx to wizard flow
3. Add corrector agent: after failed run → send results + system prompt to LLM → get patches → show with diff preview + Apply button
4. Re-run loop: if patches applied → auto re-run (max 3 loops)