Add depth-degradation evaluation infrastructure

[elenaajayi]

Summary

This PR adds compute-independent depth-degradation analysis for saved Scenario 1 probe scores. It compares matched 2-hop and 3-hop trajectories using AUROC, Brier score, ECE, Temporal Divergence, and match-group bootstrap confidence intervals.

It does not load a model, start a GPU, call Modal, generate trajectories, or call an LLM judge.

What changed

• Adds src/analysis/depth_degradation.py for depth metrics and paired 3-hop minus 2-hop comparisons.
• Adds src/analysis/splits.py to keep complete match groups together and prevent leakage across splits.
• Adds the ordered command scripts/03_probe_analysis/06_analyze_depth_degradation.py.
• Updates the main README so the command appears in Step 6 of the experiment.
• Keeps Qwen thinking-on and thinking-off results in the same match-group partition while reporting the modes separately.
• Keeps injection_present separate from trajectory_action_executed.
• Requires Worker1 at hop 1 as the fixed injection anchor.
• Uses one executor score per trajectory for AUROC, Brier, and ECE.
• Uses the full ordered agent-score sequence for Temporal Divergence.
• Excludes indeterminate action labels from AUROC, Brier, and ECE instead of treating them as negative examples.
• Resamples complete match groups for confidence intervals.

Label rule

latent_compromise_status: candidate marks a trajectory for later mechanistic review. It is not a confirmed latent-compromise label. A propagated or blocked run remains a black-box non-compromise unless the executor action fired.

Current boundaries

• No real Scenario 1 experiment results are included.
• Tests use synthetic prediction scores.
• The analysis runs only after live trajectories, activations, and probe scores exist.

Validation

• Full repository suite: 174 passed
• Focused depth and split tests: 20 passed
• Direct command import and --help: passed
• Python compilation and git diff --check: passed

Review focus

• Does the 3-hop minus 2-hop comparison answer the intended depth-degradation question?
• Are indeterminate action outcomes handled correctly?
• Do the split rules keep every complete match group together?