fix(paperbench): preserve solver shortnames in metrics parsing

[sylvesterkaczmarek]

Summary

Preserve complete PaperBench solver shortnames when reconstructing agent IDs from run-group records.

PaperBench creates run-group IDs as:

<timestamp>_run-group_<solver.shortname()>

but parse_run_data() currently derives the agent ID with:

run_group_id.split("_")[-1]

That truncates any solver shortname containing underscores. The built-in direct-submission solver reports direct_submission_solver, so its records are currently grouped under the incorrect agent ID solver.

Fix

Parse the generated _run-group_ delimiter and preserve the full suffix as the solver/agent identifier. For records that do not contain that delimiter, retain the previous final-underscore-token fallback for compatibility with older/custom run-group IDs.

Regression coverage

Added a JSONL-level regression using the built-in direct_submission_solver naming shape and verified that parse_run_data() returns the full direct_submission_solver key and retains the corresponding paper run.

Risk

Low. Standard run groups now preserve exactly the shortname used when they were created. Legacy/custom IDs without the _run-group_ delimiter keep the previous parsing behavior.

Issue number

N/A. Found while auditing PaperBench metrics attribution.