Scientific hardening evidence register

This register tracks the implementation requested in the scientific-hardening task. It starts from 53a4c07841944e5bfb40ff1b00cc32980131503b on the independent feature/seascape-scientific-hardening branch. The reviewed fcd0a83 commit is historical context; all findings are rechecked against current producers. This is an implementation record, not a claim of independent scientific validation or a completed regional release.

Baseline (before changes)

Interpreter: /tmp/seascape-roadmap-dev/bin/python, Python 3.14.6. All commands ran from this checkout unless noted. python -m pytest -q: exit 0, 410 passed, 3 skipped, 101 warnings; three skips require materialized regional artifacts. python scripts/check_docs.py: exit 0, 42 documents, 183 local links, 26 stages; external URLs were not checked. python -m pytest -q tests/test_documentation.py: exit 0, 16 passed. ruff check src tests scripts, ruff format --check src tests scripts, python -m mypy, and python -m pip check: exit 0. The offline CLI demo ran in a fresh /private/tmp workspace, exited 0 with Synthetic software acceptance: PASS (not a regional release) and emitted a Parquet table, manifest, report and two PNGs inside its owned .seascape/demo tree. An initial malformed shell invocation of the demo exited 1 before launching the CLI; the corrected command above was executed successfully. No distribution or clean-consumer baseline run yet.

Issue status

Issue	Current producer / finding	Required contract and downstream effect	Status
SC-01	geomorphic_units/build.py: _rise and _fall converted missing inputs to positive flat/narrow/near evidence; _select_labels used argmax on unsupported rows.	Missing required evidence keeps a class ineligible; all-ineligible rows have an explicit reason. Post-mapping support describes the final label. Changed labels/support require a new candidate and method identity.	Corrected in geomorphic_evidence_v2; five focused tests and the full suite (414 passed, 3 artifact-dependent skips) passed after this fix. This is software acceptance, not independent validation.
SC-02	bathymetry/build.py directly assigns pixels at each resolution; pipeline.py overwrote direct R6 counts and bands with R8-child sums while retaining direct R6 moments.	Canonical statistic family must use one declared sampling support; retain hierarchy-only parents as unavailable.	Removed the mixed-support overwrite; direct_pixel_support_v2 candidate metadata declares the change. Controlled H3 boundary, one valid plus nodata pixel, band closure, hierarchy-only nulls, and missing-source regressions are in test_bathymetry_contracts.py. Historical releases retain old semantics.
SC-03	utils/habitat_surface.py: clipping counted support cuts as habitat edges.	Habitat interfaces exclude artificial support cuts; patch identity is local to the reporting support, with truncation recorded.	Per-support geometry unions exclude boundary-coincident lines and set TOPOLOGY_QC_REASON; analytic and real H3 fixtures pass. Survey/AOI truncation beyond the reporting support still needs explicit metadata in SC-09.
SC-04	utils/habitat_aggregation.py: parent patch counts/edges were sums of child values and the largest patch was the largest child fragment; empty fragmentation was zero.	Recompute union-scale topology; no-patch fragmentation is undefined with a reason.	R6 publication now unions selected child support geometry per parent and measures inventory geometry on that union. Without supplied parent geometry, topology remains null with topology_not_computed instead of inventing values. Empty fragmentation is null with no_mapped_patch. Analytic fixtures and the full suite pass.
SC-05	benthic_substrate/classification/build.py closed modeled rock presence against source sediment percentages, encoded unsupported boulder/cobble/mixed as zero and fed an unjustified hardness index to reef suitability.	Preserve source-supported measurements separately, reject invalid percent inputs, and suppress unsupported derived quantities pending exact provider basis.	Rock is a modeled presence score, gravel/sand/mud are a separate sediment texture; only closed supported sediment shares get entropy. Unsupported classes, hard-substrate area fraction and hardness index are null with status, and reef suitability cannot recover them by reweighting. Producer regressions pass. Exact pinned December 2025 HUB Ocean product definitions were not accessible, so a joint areal closure and physical hardness calibration remain blocked, rather than fabricated.
SC-06	geomorphometry/build.py: historical PLAN_CURVATURE used the surface-tangential denominator.	Distinguish horizontal plan from surface-tangential curvature, retain the latter under its own name, and version the meaning.	PLAN_CURVATURE now uses the horizontal contour denominator and TANGENTIAL_CURVATURE preserves the former expression; geomorphometry_curvature_tpi_v2 manifest metadata and family documentation record equations/sign/scale. Analytic plane, parabola, dome, bowl, saddle, and rank-deficient tests pass.
SC-07	geomorphometry/build.py excluded flat facets from VRM because their aspect is undefined; its standardized TPI returned zero for nonzero prominence divided by zero neighbor variance. Geomorphic terrain context had the same TPI case and was corrected in SC-01.	Flat facets retain vertical normals; nonzero prominence with zero neighbor variance is undefined.	Flat VRM and mixed flat/60-degree analytic tests pass. Ring TPI in metres remains +50 in the controlled case, Z is null with a QC reason, and downstream broad Z is also null.
SC-08	Bathymetry contour distance queried nearest raster-edge crossing points and raised if no contour was present.	Segment distance, topology-preserving contours, and explicit unavailable status without dropping valid depth.	Marine-only marching squares produces interpolated segments, chooses ambiguous saddle pairings by center sign, and measures metric straight-line distance. A contour absent inside a valid crop yields null distance plus status while depth remains. Analytic line, circle, land/nodata, disconnected, and no-contour fixtures pass. Crop-edge censoring is documented; no regional accuracy claim is made.
SC-09	Biogenic habitat processing previously treated positive canopy polygons as surveyed area, used an archive-wide latest year and could promote one absent child to whole-parent absence.	Separate survey opportunity, presence, absence, observation/publication time and as-of filtering.	Explicit complete-event footprints alone establish surveyed area and implicit absence; otherwise completeness is unknown. Latest local polygon evidence and source priority determine present/absent/unknown fractions; point/line evidence remains non-areal. R6 fractions aggregate by water area. Observation and known availability years are filtered before staging; unknown availability is flagged as retrospective. Synthetic producer regressions pass.
SC-10	Source support through manifests/catalog/release/matrix: inspect sampling and uncertainty claims.	Effective source support and method identity survive R8/R6 and consumer export.	New candidate manifests carry method identity and source support. Release records, ProductArtifact, generated catalog fields and metric-matrix metadata retain those details when present; fixture tests cover resolution-specific records and missing uncertainty. The checked-in catalog describes older materializations and cannot be regenerated or represented as a new validated regional release without their source artifacts. Field definitions that remain producer-specific still require source-backed review; no independent validation is claimed.

Bounded additions and first use

Sixteen-bearing geometric fetch export records metres, search limit, censoring and source status. The optional GEBCO TID adapter validates declared release and exact raster grid alignment, counts categorical codes on the depth pixel support, and publishes depth/TID together. Deterministic fixtures pass; no regional TID source has been acquired or materialized. The capability matrix records implementation versus data/contract blocks.

Preflight and producers now use explicit selected jurisdictions. A B.C.-only water-geometry candidate consumes the existing Canadian region and territorial-zone layers; B.C.-only habitat selection disables Washington inventories. The default cross-border configuration remains intact. The real-data recipe specifies a bounded candidate command and source requirements, but was not executed: a fresh, permitted and reviewed GEBCO crop and canonical water-support inputs are not present. The new-user checklist is ready; no unfamiliar-user trial occurred. No source was acquired, no canonical product or retained release was changed, and no independent scientific validation has been performed in this branch.

Validation on the implementation branch

All commands below used /tmp/seascape-roadmap-dev/bin/python (3.14.6) in this checkout unless the command names its own executable. python -m pytest -q: exit 0, 444 passed, 3 skipped, 110 warnings; the three skips require materialized regional artifacts. The first full-suite run after the TID change exited 1 (one demo-provider provenance failure); correcting the inherited GEBCO datum/evidence claims made the rerun pass. python -m pytest -q tests/domains/environment/seascape/test_gebco_tid.py tests/test_preflight.py tests/test_water_geometry_provenance.py tests/test_consumer_acceptance.py: exit 0, 92 passed. ruff check src tests scripts, ruff format --check src tests scripts, python -m mypy, python scripts/check_docs.py, python -m pytest -q tests/test_documentation.py, and python -m pip check: exit 0; docs checked 46 documents, 198 local links and 26 stages, but not external URLs. The notebook's first sandboxed jupyter nbconvert --to notebook --execute notebooks/validation/01_TOOLKIT_VALIDATION.ipynb --ExecutePreprocessor.timeout=120 --output seascape-toolkit-validation-hardening.ipynb --output-dir /private/tmp exited 1 because loopback socket bind was denied. Repeating the same command with local-kernel permission exited 0; the executed notebook has 25 cells and no error outputs.

python -m build --outdir /private/tmp/seascape-hardening-dist-20260929 exited 1 because sandboxed isolated build dependency installation could not resolve PyPI. Repeating the isolated build with network permission and --outdir /private/tmp/seascape-hardening-isolated-dist-20260929 exited 0; the sdist and wheel passed scripts/check_distribution.py (9 required files, 5 packaged resources). The non-isolated fallback build also exited 0, but the later isolated build is the final package evidence. The first sandboxed clean-consumer and quickstart runs exited 1 at dependency installation because PyPI DNS was unavailable. With network permission the quickstart checker passed all install/demo/ plan blocks and 15 demo checks outside the checkout. The first network-enabled consumer run exited 1 at the installed-module check because its registry-count assertion still expected 75; the TID registration increased it to 77. The assertion was corrected to require both TID IDs; the final isolated-build wheel passed scripts/check_consumer_install.py: installed-runtime import, resources, CLI, demo and negative probes passed; after extras, 179 copied tests and the copied validation notebook passed. The matching isolated-build sdist passed scripts/check_quickstart.py: install/demo/plan blocks and 15 demo checks. Both ran outside the checkout and recorded reports under /private/tmp/seascape-hardening-isolated-consumer-20260929/ and /private/tmp/seascape-hardening-isolated-quickstart-20260929/. The direct local command seascape --workspace /private/tmp/seascape-hardening-demo-20260929 demo exited 0 with Synthetic software acceptance: PASS and kept its Parquet, manifest, report and two figures in the demo-owned subtree. None of these package checks is a regional build, a source acquisition, or a scientific accuracy assessment.