feat(models): add reasoning budgets and upstream prompt pruning

[3metaJun]

Summary

This adds detected-catalog reasoning-budget rewrites and ports the requested pstack prompt reductions and shipping verification-reuse exception. It depends on the priority workflow PR; its base is that PR's branch so the review diff contains only this second set of changes.

model-budget.mjs previews changes, preserves aliases and other Harness choices, and refuses writes for unresolved models or duplicate panels. Shipping reuses individual lane results only after documented build-output comparisons, while current-head CI and review remain mandatory.

Scope

• Add the budget CLI, schema validation, and six behavior tests.
• Prune repeated instructions from the affected skills and playbooks without changing provider defaults.
• Add a build-output reuse procedure and reconcile autopilot verdict wording.
• Preserve source pins and refresh reviewed target digests.

Verification

• npm test: 211 passed, 1 Unix-only test skipped on Windows.
• Both pinned-source baseline checks, strict upstream checks, package checks, and git diff --check passed.

Review risks

The prompt reductions apply to the shared skill tree. Upstream motivated them with Opus 5.5, but this PR does not establish behavioral equivalence for other models. Build-output reuse is an agent-facing procedure, not an automatic classifier. Dev-server lanes and lanes without saved build output must rerun.

---

Agent: gpt-6.1-sol via pi

[3metaJun]

Summary

This adds detected-catalog reasoning-budget rewrites and ports the requested pstack prompt reductions and shipping verification-reuse exception. It depends on the priority workflow PR; its base is that PR's branch so the review diff contains only this second set of changes.

model-budget.mjs previews changes, preserves aliases and other Harness choices, and refuses writes for unresolved models or duplicate panels. Shipping reuses individual lane results only after documented build-output comparisons, while current-head CI and review remain mandatory.

Scope

• Add the budget CLI, schema validation, and six behavior tests.
• Prune repeated instructions from the affected skills and playbooks without changing provider defaults.
• Add a build-output reuse procedure and reconcile autopilot verdict wording.
• Preserve source pins and refresh reviewed target digests.

Verification

• npm test: 211 passed, 1 Unix-only test skipped on Windows.
• Both pinned-source baseline checks, strict upstream checks, package checks, and git diff --check passed.

Review risks

The prompt reductions apply to the shared skill tree. Upstream motivated them with Opus 5.5, but this PR does not establish behavioral equivalence for other models. Build-output reuse is an agent-facing procedure, not an automatic classifier. Dev-server lanes and lanes without saved build output must rerun.

---

Agent: gpt-6.1-sol via pi