test: expand sanitized evaluation corpus

[utkarshp845]

Summary

• add three fully synthetic evaluation fixtures: generic HTTP 503 failures, mixed latency/memory-warning signals, and client-only HTTP 404s
• add the corresponding strict deterministic corpus cases
• bump the versioned corpus contract from 2026.07.1 to 2026.07.2
• document the new coverage and mark Week 6 Day 2 complete

Learning-path fit

This broadens evaluation coverage before Week 6 introduces adversarial and incorrect-confidence scenarios. It does not change the assistant rules or add provider dependencies.

Testing

• python -m pytest -q (41 passed)
• python -m ruff check app cli evals tests
• python -m evals.run_evals --json (10/10 cases, 50/50 checks, all hard gates pass)
• git diff --check

Privacy boundary

All new fixtures are synthetic. They contain no real user data, hostnames, identifiers, credentials, or copied production logs.