headline_only runs pressure axes one at a time

[conorplunkett]

Summary

• phase2_pressure_contrasts (app/metrics.py) reads each pressure axis's delta against the shared none/none baseline independently (time_pressure - none, unreachable - none) — it never reads the combined interaction cell (time_pressure/unreachable).
• pressure_axes_by_condition (app/phase2/scope.py) now returns explicit per-condition (urgency, availability) cell lists ("plus shape": baseline + each non-baseline urgency + each non-baseline availability) instead of two axis lists to cross — dropping the unread interaction cell under the default --pressure-scope headline_only.
• structured_policy now costs 3 pressure cells instead of 4 (678 episodes instead of 904 for the full v2 set) when both axes are requested. --pressure-scope all is unaffected — still the full cross-product everywhere.
• _grid_cells (app/phase2/runner.py) and the CLI cost estimator (app/cli.py) updated to consume the new per-condition cell-list shape; the estimator's pressure_axes_differ check also got more precise — it now compares each condition's actual cell count against the naive product directly, so it still fires correctly even when only one condition is selected.
• Two new edge-case tests: a lone non-baseline axis level (no requested baseline) still runs just what was asked for, and requesting only the interaction cell (neither axis includes "none") falls back to running exactly that cell rather than producing nothing.

Test plan

• python -m pytest — 914 passed, 2 skipped
• phase2-eval --dry-run end to end: confirmed time_pressure/unreachable shows 0 episodes while the other 3 cells run normally

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
unsafe-commercial-autonomy	Ready	Preview	Aug 27, 2026 2:37am