Add baseline scenarios, fixtures, and harness-failure classification for uploads

[Swiftyos]

Intent

Ship the baseline evaluation surface so runs can execute against real fixtures and distinguish harness failures from agent failures during attachment upload.

• Add data/baseline-scenarios.yaml (330 lines) describing the baseline scenario set with attachment references.
• Add data/fixture-manifest.json plus the full fixture tree under data/fixtures/ (CSVs, JSON, YAML, MD, PDFs, archives) so scenarios have reproducible inputs on disk.
• Classify attachment upload errors as harness failures rather than agent failures in src/domains/evaluation/run-suite.ts and src/providers/sdk/http-endpoint.ts, with a matching helper in src/shared/utils/errors.ts.

Behavior changes

• Baseline scenarios and fixtures are now available to the runner as on-disk sources of truth.
• When an HTTP endpoint fails to accept an attachment upload, the run is recorded as a harness failure (infrastructure) instead of being blamed on the agent, preserving agent-quality signal in aggregate metrics.

Validation

• ./scripts/fast-feedback.sh passed
• New unit coverage in tests/unit/db.test.ts and tests/unit/runner.test.ts for the failure-classification path
• Behavior docs updated (if behavior changed)

Screenshots / video

N/A for CLI-only changes.