refactor(studio): make screenshot-set review the training-truth workflow

[e54-bot]

Summary

• Replace the platform reviewed-annotation snapshot importer with a finalized screenshot-set importer (platform contract: owbastion.com#255): fetch immutable set membership by integer version, download members from prefix-scoped R2, verify size_bytes + SHA-256 + decoded layout against platform-signed metadata, resume against verified workspace files, then create a Studio batch whose human review produces labels.
• The set's accuracy mark propagates as accuracy_feedback through cases.json and review rows for prioritization only — it never becomes a transcription or training label. Studio review UI gains an "inaccurate first" ordering toggle and a 反馈有误 badge.
• Removed: snapshot import routes/helpers, import_platform_snapshot.py (replaced by import_screenshot_set.py), split.py, materialized datasets/labeled/rec/platform support, and snapshot.json/provenance.json staging in remote runners (provenance now read from the exported batch.json; only set identity travels into run metadata).
• The metadata contract forbids extra fields so QQ identity, player-account internals, grant/risk state, and evidence URLs cannot leak into a batch. Duplicate finalized-set imports are rejected with 409.

Fixes #26

Test plan

• uv run pytest -q — 301 passed (new importer contract/integrity/resume tests + Studio route tests)
• pnpm --dir training/studio/frontend build — clean
• scripts/batch_eval.py --only-run-code (CI gate) — passed
• Falsification: removing extra="forbid" fails the privacy tests; removing the SHA-256 check fails the integrity test
• Live import against a real finalized set (requires platform env + R2 prefix config)

Generated with Devin