SM1. Image eligibility, quality-control rules and anatomical-view eligibility

Source-image eligibility

The image-processing pipeline operated on project records flagged as original photographs. An image was eligible for shell detection when is_original = 'Y', its source-image quality-control status was not rejected, and a non-empty local filename was available. A rejected original or an image without an available local file therefore did not enter the detection stage. No eligibility rule based on adult versus juvenile status, shell condition, erosion or completeness was implemented.

The operational preprocessing workflow used the project YOLO segmentation weights stored as shellseg_gold_20260506_0823_yolov8n/weights/best.pt. The preprocessing script recorded the model path and SHA-256 digest, detection threshold, inference image size, crop settings and complete command arguments for every committed run. The documented full-run workflow did not override the preprocessing defaults listed in Supplementary Table SM1.

Detection and crop quality-control flags

YOLO inference was performed with a confidence threshold of 0.35 and an inference image size of 640 pixels, with high-resolution instance masks requested from the segmentation model. Predicted mask values greater than 0.5 were converted to the binary shell mask. Each valid mask was retained independently; the pipeline did not select only the highest-confidence shell when several shells were detected in the same source image.

The tight extent of each mask was expanded by 14% of the larger mask dimension, subject to a minimum of eight pixels of padding. The crop was limited by the source-image boundary, and this condition was recorded as crop_clipped. Pixels outside the shell mask were replaced by black. The resulting crop was centred on a square black canvas and reduced only when its largest dimension exceeded 800 pixels. Down-scaling used Lanczos resampling. Crop images were stored as JPEG at quality 94.

Four diagnostic attributes were calculated for each valid detection. A detection was flagged as low_confidence when its confidence was below 0.50; small_shell when its bounding-box area represented less than 20% of the source-image area; and poor_mask when the mask-to-bounding-box area ratio was below 0.15 or above 0.98. The multiple_shell flag was set whenever more than one detection was returned for the source photograph. These values were review aids, not automatic exclusion criteria.

When no bounding boxes, masks or confidence values were returned, the image was recorded with processing status no_detection and produced no downstream shell candidate. A predicted instance whose rasterized mask contained no foreground pixels likewise produced no crop. Other processing exceptions were recorded as errors rather than converted into accepted images.

Manual crop review

Generated masked crops were reviewed against their corresponding original photographs. The review interface displayed the detection confidence and the automatic low-confidence, small-shell, mask-warning and crop-clipped indicators. The reviewer could accept the crop, reject it or return it to the unreviewed state. Acceptance set the transformed image to qc_status = 'accepted' and is_analysis_ready = 1; rejection set qc_status = 'rejected' and is_analysis_ready = 0. Resetting a decision restored unreviewed status and removed analytical readiness.

Final crop acceptance was based on reviewer judgement rather than on a second fixed numerical decision rule. A free-text rejection reason could be recorded; when a crop was rejected without one, the system stored Rejected during transformation quality review. The interface also supported reviewer-initiated bulk acceptance of otherwise pending crops. Automatic detection flags therefore informed inspection but did not determine the final quality-control decision.

Anatomical-view classification and eligibility

Anatomical view was classified independently of image rotation. The reviewer was instructed to identify the shell surface presented to the camera rather than the orientation of the image on the page. The available anatomical categories and their operational definitions are given in Supplementary Table SM2.

Setting the view itself to uncertain automatically assigned uncertain confidence. A record could also be reset to unknown and unreviewed. Orientation eligibility required a dorsal or apertural view with certain or probable confidence, a stored segmentation mask, an unrejected original image and an unrejected masked crop.

The orientation program also provided an optional --require-accepted policy under which only explicitly accepted masked crops were eligible. This option was not included in the documented operational full-run command; under the standard implementation, crops were excluded when rejected rather than requiring an explicit accepted state.

SM2. Orientation geometry and construction of standardized visual representations

Automatic long-axis and apex determination

Apex-up orientation was generated by algorithm version mask_cross_v2_margin10. Geometry was measured from the original binary shell mask in source-image coordinates. Masks containing fewer than 50 foreground pixels were rejected as non-measurable. The foreground x–y coordinates were centred on their centroid, and the covariance matrix of these coordinates was eigendecomposed. The eigenvector corresponding to the largest eigenvalue defined the shell long axis. A mask whose projected long-axis length was one pixel or less was treated as degenerate.

To identify the maximum perpendicular shell width, mask pixels were projected onto the long axis and its perpendicular. Long-axis positions were discretized into one-pixel bins, and the difference between maximum and minimum perpendicular coordinates within each bin defined a width profile. This profile was smoothed using an odd moving-average window whose size was min(7, max(1, (bin_count // 20) × 2 + 1)). The bin with the largest smoothed width defined the candidate maximum-width intersection.

Pixels lying within max(1.5, kernel_size / 2) pixels of that long-axis location were used to estimate the perpendicular chord. Its endpoints were defined by the 0.5th and 99.5th percentiles of their perpendicular coordinates, limiting the influence of isolated edge pixels. If no supported perpendicular chord could be obtained, orientation processing failed rather than substituting an estimated value.

The position of the maximum-width chord along the long axis was expressed on a 0–1 scale. Because the widest shoulder region was expected to lie closer to the spire than to the elongated basal taper, a cross-position at or below 0.5 caused the first end of the PCA long axis to be assigned as the provisional apex; a value above 0.5 assigned the opposite endpoint. This rule resolved the arbitrary sign of the PCA eigenvector. The inferred apex vector was then rotated to point vertically upward.

Orientation confidence and automatic review flags

Three geometry-confidence values were calculated. Long-axis confidence was derived from the ratio of the first to second PCA eigenvalues as 1 − 1 / eigenvalue_ratio, bounded to the interval 0–1. Maximum-width confidence was the median positive width in the local neighbourhood of the selected chord divided by the selected chord length, also bounded to 0–1. Apex confidence was 2 × |cross_position − 0.5|, again bounded to 0–1. Overall confidence was the minimum of these three quantities.

Automatic orientation-review flags were generated when long-axis confidence was below 0.35 (weak_long_axis), maximum-width confidence was below 0.70 (unstable_maximum_width_chord), or apex confidence was below 0.30 (ambiguous_apex_end). The implementation-level manual-review threshold for overall confidence was 0.45; values below this threshold generated overall_confidence_below_threshold.

Three warnings from the preceding shell-detection stage were also propagated to orientation review when present: source_crop_clipped, source_poor_mask and source_low_confidence. These warnings increased review attention but did not by themselves reject an orientation.

Rotation, padding and output geometry

Before rotation, the raw RGB image and mask were cropped to the tight shell-mask bounding box. The RGB crop was masked onto a black background. Additional safety padding equal to 14% of the larger crop dimension was added around the shell, subject to a minimum of eight pixels; the implementation capped this safety padding ratio at 20%. Because this padding was created outside the original photograph rather than clipped to its boundary, a shell touching a source-image edge could still be rotated without introducing additional geometric clipping.

RGB rotation used bicubic interpolation and mask rotation used nearest-neighbour interpolation. Both rotations expanded the canvas and filled new pixels with black. The rotated shell was cropped to the non-zero mask extent and centred on a new square black canvas. The canvas dimensions were calculated so that the shell retained a minimum target margin of 10% on every side. Outputs larger than 800 × 800 pixels were reduced to 800 × 800 pixels using Lanczos resampling; smaller outputs were not enlarged. Oriented RGB images were stored as JPEG at quality 95. Horizontal reflection was never applied.

The transformation metadata stored the source-mask bounding box, number of rotation-safety pixels, rotation angle, rotated-mask bounding box, target margin, actual top, bottom, left and right margins, final output dimensions and an explicit mirrored = false flag. This matrix was subsequently reused to reconstruct the shape mask in exactly the geometry of the accepted RGB orientation.

Orientation review and manual correction

Orientation review displayed the masked source image, a diagnostic preview and the apex-up derivative. In the diagnostic preview the PCA long axis was shown in red, the maximum-width perpendicular chord in cyan and the inferred apex in green. The displayed quantitative fields included long- and short-axis lengths, relative cross position, rotation angle, apex confidence, overall confidence and the four output margins.

An automatically generated orientation without review flags initially received the internal status auto_pass; an orientation with one or more flags received pending. In either case, the generated transformed image itself was initially stored as unreviewed and is_analysis_ready = 0. Shape and pattern construction required an explicitly accepted orientation and an accepted, analysis-ready RGB derivative; an automatic geometry pass alone therefore did not qualify an image for final analysis.

The reviewer could accept, reject or reset an orientation. Acceptance marked the oriented RGB derivative as accepted and analysis-ready. Rejection removed analytical readiness and, when no note was supplied, stored Rejected during orientation quality review. A reset returned the orientation to pending status.

Manual correction allowed an angular adjustment between −180° and +180° and/or an explicit 180° apex reversal. The correction worker added the requested angle and, where selected, 180° to the original rotation and normalized the resulting angle to the interval −180° to +180°. A corrected image was written as a new derivative; the earlier orientation and transformed image were marked superseded rather than overwritten.

Reconstruction of the aligned binary shape representation

The shape image was not generated by thresholding the oriented RGB derivative. Instead, the original stored YOLO mask was reconstructed using the transformation matrix of the accepted orientation. The source mask was cropped to the recorded shell extent, padded by the same stored number of safety pixels, rotated through the same angle using nearest-neighbour interpolation, cropped to the same rotated mask extent and centred using the same 10% target-margin rule.

If the reconstructed mask dimensions differed from the stored oriented RGB image, the mask was resized to the RGB dimensions using nearest-neighbour interpolation. It was then rebinarized at an 8-bit value of 128. The resulting PNG therefore represented an aligned binary silhouette in the same coordinate system as the accepted oriented colour image.

Luminance-normalized pattern representation

Pattern normalization was implemented by algorithm version aligned_mask_lab_luminance_v1. The accepted oriented RGB image was converted to the CIE L*a*b* colour space using OpenCV [6, 11]. Calculations were restricted to pixels inside the aligned shell mask.

Within-shell values of the 8-bit LAB luminance channel were summarized by their 2nd and 98th percentiles. The lower and upper percentiles were mapped linearly to target L values of 24 and 235, with values outside the percentile interval clipped to that mapped range. The normalized luminance was then blended with the observed luminance:

Only the LAB luminance channel was replaced by this blended value; the chromatic a and b channels and image geometry were not otherwise modified before conversion back to RGB. Pixels outside the shell mask were set to black. If the difference between the 98th and 2nd within-shell luminance percentiles was less than eight L-channel units, the image was treated as unsuitable for this normalization and processing failed rather than artificially expanding the narrow luminance range.

Representation validation warnings

The reconstructed shell mask was checked before the shape and pattern images were registered as successful derivatives. Connected components were calculated using 8-connectivity. A mask generated multiple_mask_components when the number of foreground components was not exactly one, and mask_touches_canvas_edge when any foreground pixel occurred on the first or final image row or column. Foreground occupancy below 5% generated very_small_foreground; occupancy above 80% generated very_large_foreground.

Pattern normalization produced three additional warnings. The mean absolute change in within-shell L-channel value, divided by 255, generated strong_luminance_change when it exceeded 0.20. The fractions of input shell pixels at or below the lower percentile and at or above the upper percentile were also recorded; values exceeding 0.08 generated excess_low_clipping or excess_high_clipping, respectively.

These conditions set the representation validation status to warning; absence of warnings produced passed. Neither status automatically accepted or rejected the image. Shape and pattern derivatives were initially created with qc_status = 'unreviewed' and is_analysis_ready = 0.

Independent review of shape and pattern derivatives

The binary shape image and luminance-normalized pattern image were displayed separately in the representation-review interface, together with any automatic validation warnings. The reviewer could accept, reject or reset the shape and pattern components separately, or apply the same decision to both. Acceptance marked the corresponding transformed image as analysis-ready; rejection removed analysis readiness. A rejection without a supplied note was recorded as Rejected during representation quality review.

The RGB image itself was the already accepted orientation-normalized derivative and was not regenerated by the representation worker. A complete visual lineage became eligible for downstream DINOv3 extraction only when the oriented RGB image, binary shape image and normalized pattern image were all accepted and analysis-ready. Any missing, pending, rejected, superseded or failed component prevented the lineage from entering the complete three-stream cohort.

Explicit processing failures

Processing failures were retained as failures rather than repaired by undocumented fallbacks. At the detection stage these included a missing source file, complete absence of a detection or mask, and predicted masks from which no foreground crop could be constructed. At the orientation stage explicit errors were raised for a missing or malformed stored mask, an empty mask, fewer than 50 foreground pixels, a degenerate long axis, absence of valid perpendicular chords, absence of support for the selected maximum-width chord, or failure to obtain a rotated shell mask.

Representation generation failed when the stored orientation matrix lacked one of the mask-reconstruction fields, when the accepted oriented RGB image or source mask file was missing, when the stored mask was not two-dimensional or was empty, when the reconstructed mask became empty, when the RGB and mask dimensions were inconsistent, or when the within-shell luminance range was narrower than eight units. Such records received an error processing state and did not become analysis-ready representations.

Processing provenance and versioning

Each committed processing stage stored a run identifier, algorithm version, eligibility rule and serialized parameter set. Detection runs additionally stored the segmentation-model path and SHA-256 digest. Individual transformed images retained parent-image identifiers, transformation type, generation method and version, dimensions, file hash, processing status, QC status and analytical readiness. Orientation transformations recorded their geometric matrix, and shape and pattern derivatives retained the accepted orientation as their parent.

Reprocessing did not overwrite previously generated derivatives. Orientation corrections and deliberate reruns created new versioned transformed images; superseded derivatives were retained but marked non-analysis-ready. This provenance structure allowed the final analytical cohort to be restricted to complete accepted image lineages while preserving earlier processing and review decisions for audit.

SM3. Operational catalogue of physical and biological traits

Supplementary Table S1 lists all 41 predefined trait fields. Measurements were made only within the accepted shell mask or from specimen metadata. Shape traits used the binary silhouette, pattern traits used the luminance-normalized pattern representation and colour traits used the masked RGB representation. Unless otherwise indicated, proportional measurements are dimensionless and range from 0 to 1.

The catalogue contains 41 measured fields and three fields requiring future annotation. “Primary” denotes a measurement used to describe biological phenotype. “Exploratory” denotes a reproducible image descriptor whose biological interpretation has not been independently validated. “Legacy” denotes a field retained for comparison but excluded from independent evidence counts.

Measurement pooling and validity

Measurements were calculated independently for each accepted image. When several accepted images represented the same physical specimen in the same anatomical view, valid values were averaged. Mean hue was the exception and was combined by circular mean. Dorsal and apertural values were not averaged together. If an input image or shell mask was unreadable, dimensionally inconsistent or empty, measurements requiring that input were marked invalid. Missing physical measurements and invalid image-derived values were not imputed.

Interpretive limitations

Several traits share pixels, thresholds or mathematical components and are therefore correlated by construction. The number of individually supported traits was not interpreted as the number of independent biological differences. Evidence summaries emphasized non-redundant phenotype modules, effect sizes and consistency across shape, pattern and colour. Tent-like elements and rule-based colour coverages were treated as exploratory until independently validated against manual annotations.

SM4. Embedding and dimensionality-reduction implementation details

This section records the computational identifiers, artifact contracts and projection parameters associated with the embedding and dimensionality-reduction workflow. The statistical role of the embeddings, feature standardization, principal-component retention rule and restriction of nonlinear projections to exploratory visualization are described in the main Methods and are not repeated here.

SM4.1 Registered embedding-extraction contract

Embedding extraction was implemented by extract_pennaceus_dinov3.py, extractor version 2.0.1. The registered model key was dinov3_vits16_cls336_v2, using the DINOv3 preset dinov3_vit_small_lvd1689m [4]. The preprocessing-contract identifier was keras_hub_dinov3_converter_explicit336_v2.

The extraction implementation defined the following immutable model contract: TensorFlow/KerasHub framework, 336 × 336-pixel input, class-token pooling, 384 output dimensions, float32 raw vectors and a separately stored L2-normalized float32 representation. The model-weight digest encoded in the supplied extraction contract was:

With the supplied contract values, canonical serialization of the model contract gives the following contract digest:

Each committed extraction run produced an immutable artifact directory containing manifest.csv, cohort_manifest.csv, embeddings.raw.float32.npy, embeddings.l2.float32.npy, transformed_image_ids.uint64.npy, items.csv, metadata.json, result.json, run.log and SHA256SUMS. Individual embedding-vector records additionally stored SHA-256 digests for both their raw and L2-normalized byte representations.

Extraction metadata recorded the embedding run identifier and run name, representation-run identifier, host and platform, model contract, model-contract SHA-256, manifest SHA-256, embedding dimensions and dtype, together with the installed TensorFlow, Keras, KerasHub, NumPy and Pillow versions. These values remained attached to the versioned extraction run rather than being inferred later from the analysis environment.

SM4.2 Frozen input arrays used for dimensionality reduction

The dimensionality-reduction runs used the frozen reviewed specimen–view cohort:

The cohort contained 1,862 aligned specimen–view units from 1,000 physical specimens. All three streams used exactly the same ordered analysis_unit_id vector. The dimensionality-reduction worker read the following L2-normalized cohort arrays:

Before any dimensionality-reduction calculation, the worker verified the cohort's complete SHA256SUMS manifest. It then required the analysis_unit_ids.uint64.npy sequence to agree exactly with the ordered database analysis units, rejected non-finite feature values, and required the maximum absolute departure of input-vector L2 norms from unity to be no greater than 10⁻⁴.

The SHA-256 digest of each source array was taken from the verified cohort checksum manifest, or recalculated directly if not present there, and stored in Pennaceus_DimRedRun.source_array_sha256 together with the corresponding source_array_path. The supplied fixed publication pages preserve the run keys and dimensionality results but do not render the three 64-character source-array digests themselves.

SM4.3 Dimensionality-reduction run identifiers and retained dimensions

Dimensionality reduction was performed by the versioned worker compute_dimensionality_reduction_v2.py. The three fixed publication snapshots identify the following immutable run keys:

The database contract for these runs recorded sklearn.StandardScaler(with_mean=True,with_std=True) and sklearn.PCA(svd_solver=full). PCA component eigenvalues and explained-variance ratios were written to Pennaceus_PCAComponent; retained specimen–view PC scores were written to Pennaceus_PCAScore.

SM4.4 Exact exploratory projection parameters

UMAP and t-SNE were generated from the same standardized 384-dimensional specimen–view matrices used as PCA input. Their parameter dictionaries were fixed in the dimensionality-reduction worker and stored with each run.

The stored two-dimensional coordinates, together with PC1–PC3, were registered in Pennaceus_DimRedCoordinate for every analysis_unit_id. UMAP was run with a single worker (n_jobs=1) and both nonlinear methods used random_state=42.

SM4.5 Software-version provenance

Software provenance was stored at the computational stage that generated the corresponding artifacts. The embedding extractor recorded versions for TensorFlow, Keras, KerasHub, NumPy and Pillow, together with the execution host and platform. The dimensionality-reduction worker independently recorded NumPy, scikit-learn and umap-learn versions in Pennaceus_DimRedRun.software_versions_json.

The fixed HTML publication snapshots supplied with the present manuscript do not render the literal package-version strings from these registry fields. Consequently, no software version has been reconstructed from installation dates or from a different execution environment. The exact values remain attached to the corresponding immutable run records.

SM4.6 Array and database lineage

The dimensionality-reduction worker enforced the following provenance chain before writing any projection result. The requested analytical cohort had to have status frozen; its checksum manifest had to validate; the ordered analysis_unit_ids.uint64.npy array had to match the included database analysis units exactly; and the selected representation array had to satisfy its dimensionality, dtype, finiteness and L2-norm contracts.

On creation of a dimensionality-reduction run, the cohort identifier, embedding-model identifier, representation stream, sample and feature counts, scaler and PCA contracts, UMAP and t-SNE parameter JSON, software-version JSON, source-array path and source-array SHA-256 were written to Pennaceus_DimRedRun. PCA eigenvalues, retained PC scores and visualization coordinates were then linked back to the same dimred_run_id and analysis_unit_id values.

S1. Complete coverage and descriptive distributions of measured shell traits

This section provides the complete descriptive record for the 41 physical and image-derived phenotype fields used in the study. It supplements the selected distributions reported in Result 1 and does not introduce an additional inferential analysis. Operational definitions, units and measurement algorithms for the individual fields are provided in Supplementary Table S1.

S1.1 Specimen-level construction and coverage

Descriptive statistics were calculated once per physical specimen. When an apertural specimen–view was available it supplied the specimen-level measurement; a dorsal observation was used only when no apertural observation was available. The resulting descriptive dataset therefore comprised 934 apertural-preferred and 66 dorsal-fallback specimen records, for a total of 1,000 physical specimens. Dorsal and apertural measurements were not averaged together for this summary.

Thirty-eight of the 41 measured fields had complete specimen-level coverage. Stored shell length, derived shell width and physical aspect ratio were available for 837 specimens (83.7%); the remaining 163 specimens lacked the stored physical-length measurement required for those fields. The image-derived measurements and cross-intersection ratio were available for all 1,000 physical specimens. Missing values were not imputed.

Supplementary Table S1 reports the median, interquartile range and complete observed range of each field. The complete range is retained here for audit, including isolated extreme measurements. In the main Results, selected traits were instead summarized using their central 98% range when a robust description of the bulk of the distribution was more informative than the absolute endpoints. No observation was removed solely because it occupied an extreme position in one of these descriptive distributions.

S1.2 Distributional features relevant to interpretation

The complete distributions included both comparatively constrained variables and strongly asymmetric measurements. Shell outline solidity, compactness and body taper occupied relatively narrow ranges around their central values, whereas shoulder angularity, dark-fragment density, tent-like-element measurements, hue dispersion and several rule-based colour-coverages contained longer tails. Violet coverage was concentrated close to zero, while the legacy binary band-presence field was constant at zero.

These distributional properties were treated descriptively. No generic statistical outlier threshold was applied across the 41 traits, and the observed endpoints reported in Supplementary Table S1 were not automatically excluded from later analyses. Analysis-specific complete-case, estimability and robustness rules are described with the corresponding methods.

Mean surface hue requires separate interpretation because it is circular. Its median, interquartile range and observed degree range are included in Supplementary Table S1 only as convenient descriptive summaries. Inferential analyses requiring hue used the circular treatment specified in the main Methods rather than treating the degree scale as an unrestricted linear variable.

S1.3 Complete descriptive trait catalogue

S2. Complete trait–embedding associations

The relationship between directly measured shell phenotype and the image-derived embedding spaces was evaluated for the RGB, silhouette-shape and luminance-normalized pattern representations. The principal Results report only selected biologically informative associations. This supplementary section provides a complete trait-level summary across all three representations without introducing an additional statistical analysis or an additional criterion for selecting biologically important relationships.

Forty of the 41 measured fields were statistically estimable. The legacy binary band-presence field was constant and therefore had no estimable association. For each of the remaining traits, Supplementary Table S2.1 reports the retained principal component having the largest absolute partial Pearson correlation within each representation. The retained embedding spaces contained 56 RGB PCs, 19 shape PCs and 57 pattern PCs. The complete calculation therefore comprised 5,280 stored trait–PC association records (40 × [56 + 19 + 57]).

FDR q-values retain the correction scheme defined in the main Methods: correction was performed across all retained PCs within one trait and one representation. The values therefore do not represent a false-discovery-rate correction across the complete trait-by-PC matrix. Sample size refers to specimen–view units with a valid value for the corresponding trait. Physical-size measurements were available for 1,603 units; the other estimable traits represented 1,862 units.

Supplementary Table S2.1. Strongest retained embedding-PC association for every measured phenotype field in each of the RGB, silhouette-shape and luminance-normalized pattern representations. The selected PC is the component having the largest absolute partial Pearson correlation with the corresponding trait. Partial correlations are shown with their fitted sign; PCA signs are arbitrary, so sign describes direction only within the stored orientation of that PC. PC variance is the explained-variance ratio of the selected embedding PC and is not the variance in that PC explained by the measured trait. The legacy binary band-presence field was constant and therefore non-estimable.

Complete association records

Supplementary Table S2.1 is a trait-level index of the complete analysis rather than a replacement for the underlying coefficient matrix. The machine-readable supplementary data retain every estimable trait–PC combination for all three representations, including representation, trait identifier, PC index, sample size, PC explained-variance ratio, raw correlation, partial correlation, analytic probability and within-trait, within-representation FDR-adjusted q-value. This preserves associations involving lower-variance PCs that are not visible when only the strongest PC for each trait is printed.

S3. Complete trait-covariance matrix

The complete covariance structure of the 41 measured phenotype fields was examined at physical-specimen level to document relationships that are necessarily summarized only selectively in the main Results. This supplementary analysis reports the pairwise Spearman correlation structure underlying the redundancy assessment and the subsequent definition of the primary measured-trait space. It does not introduce an additional hypothesis test, correlation threshold or trait-selection rule.

S3.1 Construction of the pairwise correlation matrix

The correlation matrix used the same specimen-level phenotype records as the measured-trait structure analysis. Repeated photographs had already been pooled within physical specimen and anatomical view. For this analysis, the apertural specimen–view was selected when available and the dorsal specimen–view was used as fallback. Stored shell length was treated as specimen metadata and was averaged across retained views where required. Missing measurements were not imputed.

Spearman rank correlations were calculated independently for every pair of measured fields using all physical specimens having finite values for both members of that pair. Average ranks were assigned to tied observations. Consequently, the matrix was pairwise complete rather than restricted to the 837-specimen complete-case population used for the primary trait PCA. Correlations involving the physical-size fields could contain 837 specimens, whereas pairs among fields with complete coverage generally contained all 1,000 physical specimens.

The legacy binary band-presence field was constant in the frozen trait run. Pairwise correlation with a constant variable is not estimable and is therefore represented as unavailable rather than as a correlation of zero. The diagonal otherwise represents the usual self-correlation of a variable.

S3.2 Complete correlation matrix

Supplementary Table S3 contains the complete pairwise Spearman correlation matrix for the 41 measured fields. Cell values are Spearman ρ. Pairwise sample size is determined by the valid observations available for the two corresponding traits and therefore is not necessarily identical throughout the matrix.

1. T01 Median dark-fragment size median_dark_fragment_area
2. T02 Mean relative tent size mean_tent_relative_area
3. T03 Median relative tent size median_tent_relative_area
4. T04 Stored shell length shell_length_mm
5. T05 Stored shell width shell_width_mm
6. T06 Outline compactness outline_compactness
7. T07 Shell slenderness slenderness_ratio
8. T08 Physical aspect ratio physical_aspect_ratio
9. T09 Outline solidity outline_solidity
10. T10 Shoulder angularity shoulder_angularity
11. T11 Body taper ratio body_taper_ratio
12. T12 Shoulder width ratio shoulder_width_ratio
13. T13 Spire included angle spire_angle_deg
14. T14 Relative spire height relative_spire_height
15. T15 Maximum-width position max_width_position_ratio
16. T16 Legacy spire-height field legacy_spire_height_ratio
17. T17 Cross Intersection Ratio cross_position_ratio
18. T18 Hue dispersion hue_dispersion
19. T19 Orange coverage orange_coverage
20. T20 Mean surface hue mean_hue
21. T21 Violet coverage violet_coverage
22. T22 Mean chroma mean_chroma
23. T23 Mean saturation mean_saturation
24. T24 Brown coverage brown_coverage
25. T25 Lower-third pattern density lower_pattern_density
26. T26 White coverage white_pattern_coverage
27. T27 Mean surface lightness mean_lightness
28. T28 Pattern density pattern_density
29. T29 Middle-third pattern density middle_pattern_density
30. T30 Upper-third pattern density upper_pattern_density
31. T31 Dark-fragment density pattern_fragment_density
32. T32 Tent-like element density tent_density
33. T33 Reticulation edge density reticulation_edge_density
34. T34 Tent-like element count tent_count
35. T35 Detected band count band_count
36. T36 Pattern luminance entropy pattern_entropy
37. T37 Colour heterogeneity colour_heterogeneity
38. T38 Surface colour contrast colour_contrast
39. T39 Continuous band strength band_strength
40. T40 Outline asymmetry outline_asymmetry
41. T41 Legacy binary band presence legacy_band_presence

S3.3 Exact and near-exact redundancy

The complete matrix was interpreted together with the implemented trait definitions. Three relationships identified duplicated or effectively duplicated measurements, while one additional legacy field was non-variable. These cases are summarized below because they determine why the corresponding fields were not allowed to contribute independent dimensions to the primary trait space.

S3.4 Covariance beyond mathematical duplication

Strong covariance was also present among measurements that were retained as biologically interpretable traits. Such relationships were not classified as redundancy merely because their correlation was large. The frozen specimen-level audit showed, for example, close covariance among overall and regional pattern-density measurements, inverse covariance between pattern density and shell lightness, covariance among related spire measurements, and covariance among related colour-intensity measurements. These relationships represent shared phenotype structure and were retained rather than eliminated solely to reduce correlation among the primary variables.

The distinction is important because the primary phenotype space was intended to describe coordinated morphology rather than to construct an artificially orthogonal catalogue of biological characters. Exact or effectively duplicate measurements were prevented from contributing repeatedly, whereas covariance among distinct primary traits remained available to the subsequent multivariate analysis.

S3.5 Clustered visualization of the complete matrix

For visualization, the 40 estimable traits were ordered by deterministic average-linkage hierarchical clustering using the dissimilarity 1 − |ρ|. Thus, both strong positive and strong inverse relationships contributed to proximity in the display. At each merge, leaf orientation was chosen to minimize adjacent dissimilarity, with trait keys resolving exact ties. The constant legacy band-presence field could not enter the clustering and was placed last. The same ordering was applied to rows and columns. This display is intended only to make blocks of positively or negatively covarying traits easier to inspect; it did not define phenotype modules, select traits or constitute evidence for discrete biological clusters.