fix: gate intensity state on CLI, writers, and JOSS docs

[D-sudoasd]

Reviewer report (file-backed review of main @ 3e599dd)

Verified each finding against 3e599dd, then fixed what a coding agent can fix. Author metadata, research-use case, public-history date, and v2.0.0 tagging were left untouched. estimate_k_factor_robust math and propagate_absolute_uncertainty were not churned.

1. Intensity-state gates were not on the public contract — confirmed on main, fixed

On 3e599dd, cli.py estimate-k never called require_relative_input_for_absolute_scaling. tests/test_cli.py::test_cli_estimate_k_accepts_scattering_column_names_with_units fed I_abs and recovered K=2. subtract_buffer had no state check (docstring said “cm⁻¹ or relative”). minimal_2d applied K/d with no state object.

Now: CLI estimate-k gates relative input and refuses unlabeled I and I_abs columns. New subtract-buffer and the kernel require absolute sample/buffer provenance. minimal_2d gates the standard as relative before K (K-only after thickness is already on the ledger). Paper/README no longer claim a universal contract for the thin utilities (norm-factor, parse-header, parse-external1d).

2. Writers stamped 1/cm; readers always returned i_rel — confirmed on main, fixed

On 3e599dd, writers hardcoded unit="1/cm". Readers always set i_rel/err_rel. tests/test_io_formats.py wrote i_abs = 100/q and asserted result["i_rel"].

Now: writers stamp 1/cm only for absolute_cm^-1 plus an explicit cm⁻¹ unit. Readers always return intensity/uncertainty and add i_abs or i_rel only for the assessed state. Tests assert i_abs and refuse unlabeled writes.

3. Paper claimed XSD + punx as software facts — confirmed on main, docs downgraded

CI still does not run xmlschema/punx. Paper/README/FAQ now say: offline 2026-08-15 check, not in CI; no third-party consumer validation.

4. Review unreleased 2.0.0 on main, not v1.1.1 — fixed in docs

README, paper, FAQ, and readiness notes say to review the unreleased 2.0.0 tree on main. No tag or release was created.

5. CLI had no --thickness and accepted unlabeled I — confirmed on main, fixed

--thickness-cm is centimetres. Help text keeps Workbench Tab 1 in mm (1.055 mm = 0.1055 cm). Unlabeled I is refused.

6. minimal_2d overclaimed — confirmed, tightened; one clause discarded

The example is a 9×9 homemade integer-bin radial average, not pyFAI, with unknown uncertainty. Paper, README, example README, FAQ, and test_minimal_2d_example now assert detector_shape=[9,9], reduction=homemade_integer_bin_radial_average, uses_pyfai=false, uses_bl19b2_campaign=false.

Discarded clause: “NIST path bypassed” is only true for the BL19B2 campaign runner. The example still calls build_nist_net_image for detector-space NIST-blank subtraction. Docs now say that explicitly.

7. Unitless "absolute" mapped to cm⁻¹ — confirmed on main, fixed

_state_from_metadata no longer aliases bare "absolute" to ABSOLUTE_CM_INV. That label is ambiguous. Writers refuse it.

What changed on this branch

• Intensity-state gates on CLI estimate-k, CLI subtract-buffer, subtract_buffer, and minimal_2d
• Conditional 1/cm writers; state-keyed readers; tests no longer enshrine i_rel for absolute data
• --thickness-cm; refuse unlabeled I and I_abs columns
• Docs: review 2.0.0 on main; XSD/punx offline 2026-08-15; minimal_2d limited to the 9×9 homemade reducer

Tests run

• pytest -q — 734 passed
• ruff check SASAbs.py saxs_mpl_style.py src tests paper/*.py scripts/*.py — all checks passed
• python examples/minimal_2d/run_minimal_2d_pipeline.py — planted K recovered (k_relative_error ≈ 0.00193)
• Pandoc is not in this environment, so scripts/check_submission_readiness.py was not rerun

Author-only leftovers (not invented)

• Corresponding author, current email, author list/order, ORCID confirmation, CRediT
• Funding, acknowledgements, competing-interest statements
• Recoverable AI product/model/version details
• A verifiable research-use case
• Six-month public-history gate: wait until 2026-08-26
• Do not create v2.0.0, a GitHub Release, or a Zenodo version archive