Complete Prime research claim evidence

[intertwine]

Summary

• add Prime 0.6.2 hosted-training workflow support, model/run planning, and collection scripts
• attach completed Prime hosted run manifests/results for the SV-Bench research claim run set
• update leaderboard, training index, and technical report to distinguish E2 matched claim-grade evidence from E1 exploratory evidence
• harden E2 reward/tool evidence handling, fallback scanner provenance, and reward-source comparison validation

Validation

• make check
• make svbench-v0.1-check
• make suite-v1-check
• uv lock --check && git diff --check
• uv run python scripts/prime_lab_check.py
• uv run svbench_manifest validate results/runs/svbench-research-claim-20260515T141203Z/*/run_manifest.json
• uv run svbench_compare_rewards --env e2 --executable results/runs/svbench-research-claim-20260515T141203Z/hu1j9b61il8dydv2bh1aj97d/run_manifest.json --judge results/runs/svbench-research-claim-20260515T141203Z/h4y9hb6fjom4xfbclitb2938/run_manifest.json --hybrid results/runs/svbench-research-claim-20260515T141203Z/hzu95jfm266370x9kr9wwgkz/run_manifest.json --out /tmp/e2_reward_comparison_check.md

Notes

• Repo-local Prime is authenticated and hosted-ready at 0.6.2; /Users/bryanyoung/.local/bin/prime still reports 0.5.37 and was not used for claim runs.
• Secondary review findings were addressed before commit: tracked manifests now point at tracked results artifacts, leaderboard schema includes evidence grade/matched claim fields, and E2 scanner backend provenance is documented/emitted for future runs.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: da360c8b17

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve Terraform type before primary-tool filtering

For Hub/build E2 examples, _convert_e2_format() serializes only info into answer, while the dataset schema stores the language under meta.lang; those Terraform answers therefore reach _oracle_from_answer() without fixture_type and default to k8s. When a correct Terraform completion and oracle both use semgrep/... IDs, this line selects the kube-linter/ prefix and filters both sides to empty, so executable and tool-observation rewards become 0 for the Terraform split. Please carry meta.lang/fixture_type into the answer or infer the primary tool from the IDs before filtering.

Useful? React with 👍 / 👎.