v0.19.2 — CI compatibility fix

Compatibility patch for the public alpha. Use this release instead of v0.19.1 for cross-platform source validation.

• Fix a test that assumed every Python build rejects deeply nested JSON at the same decoder depth. Both decoder rejection and schema rejection remain valid fail-closed outcomes.
• Add a deterministic check that decoder recursion errors return a machine-readable CLI error, exit code 2, and false input-contract/training-readiness flags.
• Rebuild the source ZIP, English materials bundle, sdist, and wheel as version 0.19.2. Refresh the changelog and published validation record.

Validation: the local Python 3.12 suite ran 712 tests successfully with 11 optional checks skipped. The extracted-source suite, 443-file public manifest, materials audit, distribution checks, installed-wheel checks, fixture validation, and offline demo passed. GitHub CI checks Python 3.10 and 3.12.

All six attached assets include the corrected source or its checksums. Runtime input validation is unchanged. This remains an infrastructure research prototype with English documentation and an MIT license; it does not introduce forecasting model results, GPU training, or a hosted leaderboard.

v0.19.1 — Public alpha research release

Superseded by v0.19.2. The cross-platform CI test fix is included in all v0.19.2 source and materials archives. Please use v0.19.2.

Initial public alpha of Future Prediction Bench: probability forecasting protocols, delayed-reward rollout preparation, and verifiable software task environments.

• MIT license, maintainer metadata, third-party notices, and repository links.
• English documentation with a five-minute offline demo and detailed research records.
• Public packaging works from a clean checkout without the optional Chinese briefing.
• Forecasting adapters cover MLB and USGS; software environments include Docker and experimental QEMU backends.

Local validation: 711 tests ran successfully with 11 optional checks skipped. The 443-file source archive, materials audit, distribution builds, installed-wheel checks, and offline demo passed.

This is an infrastructure research prototype. It includes scripted integration evidence, but no formal forecasting model results, hosted leaderboard, or complete GPU training pipeline. Text trajectory exports remain trainer_ready=false.

See README.md, CHANGELOG.md, and docs/VALIDATION.md for usage and scope.

CI compatibility note: The initial Linux run exposed a Python JSON decoder depth assertion that varies by build. The assertion has been corrected on main; input rejection behavior is unchanged. Published v0.19.1 archives retain the original assertion. Use v0.19.2 for cross-platform test execution.