[DO NOT MERGE] exp #651 E2: digested trajectories

[eranra]

Experiment E2 of #651 (closes the run for #654 when filled in). Do not merge. This branch only adds run configuration on top of #657:

• runs-on: [self-hosted, ibm-vpc, skillberry-1], because RITS is reachable only there;
• in full_verified/overrides.env: SB_EMPTY_SEED=1 and SB_REUSE_LATEST_BASELINE=1, the same overrides as baseline 37204406089;
• CAPEVOLVE_OPTIMIZER_PROFILE=digest_trajectories.

Dispatch:

gh workflow run benchmarks.yml --ref exp/651-E2 -f benchmark=spreadsheetbench -f tier=full_verified -f iterations=5 \
  -f agent_model=ibm-rits/google/gemma-4-31B-it -f optimizer_model=ibm-rits/google/gemma-4-31B-it \
  -f gate_k_se=0.2 -f algorithm=hill-climb-all

The measures and the decision rule are pre-registered in #651 and #654. The result will be posted on #654.

🤖 Generated with Claude Code

[skillberry-bot]

🏷️ Automatic Labeling

I've analyzed this pull request and added the following labels:

• paper-experiment - algorithm - benchmark-full - paper-experiment - algorithm - benchmark-full - wontfix

These labels were selected based on the PR title, description, and changed files. If you believe any labels are incorrect or missing, feel free to adjust them manually.