Cycle 4 sticky-state action-symmetry Vast battery (5 seeds × 3 variants)

[Al-does]

Summary

• Adds run_seeds.sh for sequential five-seed Vast launches per variant.
• Recovers all 15 compact result trees (seeds 42–46 × variants 1–3) from B2 into the repo.
• Adds multi_seed_summary.json and findings.md comparing cycle 4 (sticky baseline s2=[0.30,0.30,0.40]) to cycle 2.

Key findings

variant	mean final MSE	vs cycle 2
1	0.00604	worse (+0.0035)
2	0.00350	better (−0.0015)
3	0.00121	slightly better (−0.0001)

Ranking unchanged: v3 ≪ v2 < v1. Variant 2 now maintains a stable ~33/67 noop/positive mix instead of collapsing to always-positive.

Launch details

• Vast instances: 46223593 (v1), 46223851 (v2), 46224326 (v3)
• Experiment ref: 8cd58c4, harness ref: 38f5136 (custom transition_matrix in control_model)
• Full checkpoints remain in B2; compact results committed under variant_*/results/mess3-rsa-c4-v*-seed*/

Test plan

• Local smoke test for variant_1 before launch
• All 15 runs completed on Vast with B2 upload confirmed
• Compact results recovered from B2 and committed