Add Qwen 3.6 Plus/Flash, 3.5 Plus/Flash snapshots, and Qwen3.6 27B benchmark results

[SkalskiP]

Summary

• Adds five Qwen models via OpenRouter: qwen-3.6-plus, qwen-3.6-flash, qwen-3.5-plus-20260420, qwen-3.5-flash-02-23, and qwen-3.6-27b, each with full-dataset runs for all six tasks at effort low plus reasoning at effort high (35 result files).
• Adds a --concurrency option to vlm-exam run (thread pool, order-preserving, thread-local image-encode cache in the OpenRouter provider) and extends --resume-file to also re-run empty-content samples via the new is_incomplete_sample helper, with tests.
• Documents the results publishing policy in AGENTS.md: runs are committed as they came out; the only allowed cleanup is one resume pass for self-inflicted failures (e.g. 429s caused by --concurrency).
• Regenerates web/benchmark_summary.json and all leaderboard/efficiency charts.

Detection coordinate format

All five models use detection_coordinate_format: xyxy_normalized_0_to_1000, confirmed by a 20-image local format probe on each model before the full runs (the format matching Qwen's documented bbox_2d-style absolute output scored near zero for these API models, while 0-1000-normalized xyxy parsed and scored consistently). Pricing per model comes from the OpenRouter model pages (openrouter.ai/qwen/qwen3.6-plus, qwen3.6-flash, qwen3.5-plus-20260420, qwen3.5-flash-02-23, qwen3.6-27b).

Run provenance

• All low-effort runs and the qwen-3.6-flash high run are single-pass.
• qwen-3.5-plus-20260420 high: one resume pass over 25 rate-limit (429) errors caused by our own concurrent request rate.
• qwen-3.6-27b high: parallel run with one resume pass over empty-content samples; 1 empty remains and is counted as wrong, consistent with existing leaderboard entries (e.g. qwen-3.5-27b high has 7 empties).
• No existing model's committed results were modified.

Known follow-ups (not in this PR)

• A uniform retry-failures pass across all committed reasoning runs (partially prepared locally) if we decide to change the methodology.
• Incomplete samples in some committed non-reasoning runs (e.g. kimi-k2.6 detection has 18 empties).

Test plan

• ruff check and ruff format --check pass on changed files
• Full pytest suite passes on the branch (209 tests)
• web/benchmark_summary.json regenerated with the detection dataset; five new models present at both efforts, no unrelated models included