Exp 013: Summary growth dynamics (#119)

[jimmytacks]

Summary

• Ran 3 sweeps (84 total simulation runs) testing logarithmic vs fixed summary growth across all strategies, coefficient sensitivity, and interval interaction
• Strategy rankings completely stable — lcm-subagent wins at all growth models, coefficients, and intervals tested
• lcm-subagent advantage amplified from 8.2% to 12.1% over incremental at 200 cycles under logarithmic growth
• 30k interval recommendation validated and strengthened — under logarithmic growth, 15k becomes 24-28% more expensive for incremental (double artefact eliminated)
• Phase 1-3 conclusions are fully robust under more realistic summary growth modelling

Closes #119

Test plan

• All three sweeps completed successfully with valid output
• Results consistent with prior experiment baselines (fixed-growth numbers match Exp 003/008)
• FINDINGS.md updated with new section, updated limitation UI: context size, cost breakdown, and cache hit rate charts #7, updated experiment index and programme status
• Journal entry complete with hypothesis evaluation, quantitative results, and limitations

🤖 Generated with Claude Code