Document QR v2 conditioning hardening

[msaroufim]

Summary

• Add a QR v2 README documenting the conditioning-hardening pattern for mixed batches, ranked stress distributions, and hidden seed variation.
• Clarify that tests: is a local smoke path and that competition-enforced numerical robustness belongs in benchmarks:.
• Add ranked 512 x 512 rowscale and nearcollinear benchmark cases so shape-only precision routing is checked in leaderboard mode, not only local test mode.

Why

QR v2 allows implementations to choose internal precision strategies, but the benchmark should not over-reward routing based only on public shape IDs. The important guardrail has to live in benchmarks: because leaderboard mode validates benchmark outputs and ranks those cases. tests: can still help local iteration, but it should not be treated as the enforcement mechanism.

Validation

• ruby -e 'require "yaml"; doc = YAML.load_file("problems/linalg/qr_v2/task.yml"); puts "tests=#{doc["tests"].length} benchmarks=#{doc["benchmarks"].length}"; raise "missing README" unless File.exist?("problems/linalg/qr_v2/README.md")'
• uv run python - <<'PY' ... checked inline task specs and QR case names ... PY
• git diff --check

Not run: QR numerical checker with torch, because the local Python/uv environments available in this workspace do not have torch installed.