Add multi-seed robustness study

[almondsun]

Summary

• run the fixed ByteBPE512 early-stopping protocol across preregistered seeds 1337, 2027, and 4242
• add a bounded aggregation CLI that verifies a canonical experiment fingerprint before comparing runs
• report descriptive mean, population standard deviation, range, stop-step variation, and controlled generation diagnostics
• harden run-summary/config parsing and training-seed validation

Results

• best BPC: 2.022544 ± 0.012358 population SD
• range: 2.008269–2.038414; all seeds beat the corrected character control 2.075981
• best step range: 1750–2250; stop step range: 2500–3000
• seed spread is about 49× milestone 022’s apparent weight-decay gain, confirming that delta was not decision-grade

Validation

• uv run --frozen --extra dev make check (141 tests, 91.41% coverage)
• uv run --frozen --extra dev make audit
• two additional controlled training runs and best/final generation diagnostics
• verified experiment fingerprint ed9650e463830d0d85dbbc32a40164b6699d7033017386a2cb22f2ab035431b0
• correctness and security audits completed with all findings resolved

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.