Improve Codex Spark audit reliability and tighten report prompt budget

[anand-testcompare]

Summary

• Prioritize sandbox-local OpenCode binaries (~/.bun/bin/opencode, then ~/.local/bin/opencode) before PATH lookup so sandbox runs use the newest model catalog and avoid stale global installs.
• Tighten the audit prompt execution budget and section requirements to keep generated findings concise, fragment-friendly, and more likely to complete within bounded tool usage.
• Add explicit guardrails that require writing the report and stopping immediately, reducing handoff/loop behavior that previously caused missing findings files.

Validation

• bun run check
• bun run typecheck
• bun run analyze -- --out-dir findings-codexspark53-xhigh-single2 --model "openai/gpt-5.3-codex-spark" --variant xhigh "https://github.com/ben-vargas/ai-sdk-provider-gemini-cli"
• bun run analyze -- --out-dir findings-codexspark53-xhigh-rerun-final --model "openai/gpt-5.3-codex-spark" --variant xhigh "https://github.com/ben-vargas/ai-sdk-provider-gemini-cli" "https://github.com/pablof7z/ai-sdk-provider-codex-app-server" "https://github.com/BjornMelin/codex-toolloop" "https://github.com/bear2u/codex-app-server-example"
• bun run analyze -- --out-dir findings-codexspark53-xhigh-rerun-final-retry --model "openai/gpt-5.3-codex-spark" --variant xhigh "https://github.com/BjornMelin/codex-toolloop" "https://github.com/bear2u/codex-app-server-example"
• bun run analyze -- --out-dir findings-codexspark53-xhigh-rerun-bear2u-retry2 --model "openai/gpt-5.3-codex-spark" --variant xhigh "https://github.com/bear2u/codex-app-server-example"