[Automated AI Experiment, Ignore]

[macromackie]

Summary

GOTA repeatedly copies unit paths during read-only scans and sorts the same observations for teammates. Reusing frozen team observations, borrowing scan data, and moving simulation state removes that repeated work. Visibility caching now transfers existing ray lists with one lookup, and the Wasm build enables link-time optimization.

Replay export also made a second full payload copy, exhausting the 96 MiB Wasm cap for one real input during validation. Appending directly to the output buffer removes that copy. Exhausted allocations now trap explicitly instead of risking corrupted output.

Game rules, policy budgets, decision order, scoring, and replay format stay unchanged.

Manual Tests

• Twelve frozen real inputs, six league and six experience, passed 24 local and 24 Cloudflare executions. Results and replay hashes matched the frozen references. Private diagnostics matched across variants and environments.
• The ten hosted pairs retaining the same DO constructor averaged 76.4s before and 51.1s after, a 33.1% runtime reduction. Two pairs with object recreation were excluded from this strict comparison.
• Four additional seed/roster variants passed eight local executions. Two diagnostic AddressSanitizer runs completed without errors.
• Observation cache regression assertions were mutation-tested. A bounded allocation check confirmed an explicit OOM trap.
• Peak optimized linear memory reached 92.6 MiB. Broader memory-headroom testing and full executor lifecycle checks remain adoption gates; this is not a production rollout.
• Native CI on this commit was dispatched before publication. Linux and macOS passed; Windows is still running. Registering the native GitHub stack also triggered PR checks on the same head.

PR Stack

• Optimization layer above [Automated Codex Experiment, Ignore] #77, the headless Wasm runner. Both remain drafts.

[Prepared with Codex]