[KARLalpha4768]

• CORR-01 conjunction_screener: rotate each RTN covariance to ECI before combining.
• CORR-02 ascom_hal: f-string with escaped quotes (SyntaxError before Python 3.12) rewritten.
• CORR-03 non_gaussian_pc: replace inertial misuse of CW equations with RK4 integration of two-body+J2 variational equations.
• CORR-04 bayesian_iod: admissible range bounds from ray/altitude-shell intersection.
• CORR-05 thermospheric_model: enable the MLP only when trained weights load (DENSITY_MODEL_PATH), else empirical corrections.

No unit tests currently target these code paths; numerical review recommended.

Verified (outside the Codex sandbox, Python 3.13.12 on Windows): python -m py_compile on every changed .py and node --check on every changed .js on this branch: 0 failures. pytest -q: 25 passed (baseline before changes: 25 passed). Secret scan: no hardcoded demo API key, Cesium-style JWT token, or postgres credentials remain once all five PRs are merged (each branch alone only removes its own items).

Not verified: docker build (Docker not available on the dev PC), terraform validate/plan, runtime/deployment tests. The existing tests do not exercise most changed code. Changes drafted with headless Codex and reviewed by hand.

One of five PRs splitting the HIGH findings from antigravity-findings-v2: fix/secrets-env, fix/api-nlp-lockdown, fix/docker-build, fix/infra-hardening, fix/correctness-high.