fix(db): include techniques in result-row uniqueness

[aistackdev]

Summary

PR #1 left benchmark_results_unique and eval_results_unique unchanged when it demoted spec_method to a per-measurement techniques jsonb. With two MTP variants (mtp_layers=4 vs 6) now sharing one config_id, rows with the same (workflow_run, config, isl, osl, conc) but different techniques collide on the existing constraint and the later insert overwrites the earlier one.

First post-deploy ingest of gemma4 run 26387431571 landed 6 of ~24 expected rows for exactly this reason. The remaining 18 silently got clobbered.

Migration 007 adds techniques to both unique constraints and re-truncates so re-ingest from scratch is clean. The bulk insert paths also need techniques in their ON CONFLICT clause and in their in-batch dedup discriminator — fixed in benchmark-ingest.ts and eval-ingest.ts.

Verification (post-merge)

# Re-dispatch the same run after the auto-deploy
gh workflow run auto-ingest.yml -R vngcloud/InferenceX-app -f run_url=https://github.com/vngcloud/InferenceX/actions/runs/26387431571

# Check: total bmk rows should be ~24 (was 6); distinct techniques should be 2
ssh -p 234 hoanq333@61.28.228.19 'docker compose -f /opt/docker-compose/docker-compose.yml exec -T postgres psql -U inferencex -d inferencex -tAc "
SELECT count(*) FROM benchmark_results;
SELECT techniques::text, count(*) FROM benchmark_results GROUP BY techniques;"'

Test plan

• typecheck / lint / fmt clean
• Post-merge: re-ingest of gemma4 lands ~24 rows, with both mtp_layers=4 and mtp_layers=6 represented

🤖 Generated with Claude Code