[Lulzx]

Summary

• expose --calibration-manifest on the public cactus convert command and forward it through local CQ conversion
• make cached output fail clearly instead of silently ignoring a supplied calibration manifest
• report zero-sample and unusable-Hessian RTN fallbacks separately, with strict-mode failures when GPTQ calibration cannot be applied
• document the manifest format and add CLI, report, malformed-Hessian, and tiny Needle end-to-end calibration coverage

Root cause

The internal converter already accepted calibration manifests, but the public CLI did not expose or forward the option. This made finetuned Needle conversions run without representative activation data. In addition, cached bundles could bypass calibration silently, and positive-sample Hessians that could not be applied fell back to RTN without a visible warning.

Impact

Finetuned checkpoints can now supply representative language data through the supported CLI. Conversion summaries distinguish successful GPTQ calibration from both zero-sample and unusable-Hessian fallbacks, making accidental RTN conversion visible and actionable.

Validation

• PYTHONPATH=python python3 -m pytest python/tests/test_cli_calibration.py python/cactus/convert/tests -q (77 passed)
• focused Ruff checks passed, excluding three pre-existing violations in untouched lines (E401, F401, F841)
• python3 -m compileall -q python/cactus/cli python/cactus/convert python/tests/test_cli_calibration.py
• git diff --check

Fixes #767