docs(design): define DeepSWE benchmark campaign

[fazpu]

Summary

• record what DeepSWE can currently score and publish, and why a multi-model loopy-loop result is an agent-system result rather than a model row
• define four benchmark arms (M0, H1, L1, L3) that separate the fixed DeepSWE scaffold, direct team-harness, loopy-loop orchestration, and heterogeneous model routing
• specify the Pier adapter, frozen benchmark workflow, timeout semantics, committed-patch hygiene, total usage accounting, artifact set, failure policy, and phased execution plan
• preserve the existing D1-D7 architecture boundaries; this PR changes documentation only

Important boundary

The benchmark design is accepted for this repository, but implementation remains pending. Before a full campaign, Datacurve still needs to confirm the submission category, resource limits, repetitions, and artifact/accounting contract.

Validation

• uv run pytest — 202 passed
• uv run ruff check .
• uv run pyright — 0 errors, 0 warnings
• relative documentation links resolve
• git diff --cached --check