nordisk/facett

facett de-duplication campaign — record

Class: history. The running record of the duplication-audit (clusters A–L) consolidation: what was folded to a single home, and — just as important — what was left deliberately because a merge would ADD code, change output, or carry real regression risk for no true-dup gain. Descriptive "how the shared home works" lives in the relevant *-design.md; this file is the ledger + the honest "why not" for the clusters left standing.

The audit (duplication-audit-2026-07-28) ranked twelve true-dup clusters A–L across facett's ~70 crates. Consolidation ran in passes; a later adversarial verifier confirmed the work was real but that cluster J had two byte-identical survivors and that A/D/E/G/H/I were still open. This is the settled state.

FOLDED TO ONE HOME (true dups, done)

• B — warm-red / amber literal (~24×) → facett_core::look::{RED, AMBER} (look/color.rs). facett_stat::Verdict::color re-exports them, so there is one warm-red in the workspace. No 224,90,90 / 224,72,72 Color32 literal remains outside core (the facett-demo widget-gallery [224,90,90,255] is a raw palette swatch array being displayed, not a use of the colour). Done in eb28853.
• F — lerp_color (4×) + heat_color (2×) → facett_core::look::{lerp_color, heat_color}. Done in eb28853.
• C — terrain DEM (~340 LOC map3d ↔ flight) → lives once in facett-map3d; flight pub uses it. Done in 1cd703f.
• J — serial FR step + box-fit → facett_core::barnes_hut::{fr_step_serial, normalize_box} (4fcf438), plus the two survivors the verifier found (197aab5): the PRIMARY 3-D facett-graph3d Layout3D::Force inline loop now delegates fr_step::<3> / fr_step_serial::<3> (temp 0.04), and facett-graphview::gpu_layout::normalize_box deletes its byte-identical local copy for core's. See force-layout-twolane-design.md §f2b (incl. the coincident-node behavior note — left as-is, documented there).
• L — Sutherland–Hodgman rect clip → facett_core::render::cpu::scissor (9c302b4).
• G — status chip (RichText form) → facett_core::look::chip. facett-deploykey and facett-buildthing::deck had byte-identical local fn chip; both now delegate (7c68d57).
• I — map fill-way paint recipe (4×) → facett_map::tessellate::paint_filled_way (takes already-projected ring + holes; fill: Option so MapFacet strokes unfilled polygons; outline_holes flag). facett-map MapView + MapFacet and facett-geomap GeoMap + osm2d delegate; each keeps its own genuinely-different projection preamble (f32 to_px / f64 cam.to_px + S-H clip / tilt-fade). Byte-exact: the rabbit_2d / real_map_2d / basemap PNG goldens are unchanged (7c68d57).

LEFT DELIBERATELY (would add code / change output / real regression risk)

• A — camera / pan-zoom (~10 copies). The single biggest structural item, and the one deliberately deferred. facett-core already carries the intended homes: the 2-D nav::Navigable and the 3-D render::camera::Camera/V3/Projected/view_proj, which is reproduced bit-for-bit from facett-map3d's and already goldened for equivalence by the camera_seam test (2-D and 3-D). So the seam exists and is proven; what remains is a typed migration, not a byte-safe swap:

  • The 3-D finish would replace facett-map3d::camera's own V3/Projected/view_proj (map3d adds view_proj_mirrored water + the flight floating-origin camera on top) with core's types, rippling through map3d and facett-map3d-flight (which re-exports map3d) and their 3-D render goldens.
  • The 2-D finish would move graphpan/graphview/map's cameras onto Navigable, but those copies have diverged by coordinate space and feature (map carries bearing_deg+tilt_deg in f64; the graph cameras are f32 pan/zoom with a clamp(0.25,4.0)), so a single adopter is a behavioral rewrite across panes that all carry snapshot goldens. Verdict: DEFERRED — the duplication is intentional-and-goldened; completing it is a multi-crate typed/behavioral migration whose risk is not justified by dedup alone.

• D — graphpan ≈ graphview (whole-component). The GraphNode/GraphEdge/GraphModel dup is real, but this is a crate merge: facett-graphpan is a distinct Facet (GraphPanView) depended on by facett-graphnav, facett-buildthing, facett-demo, facett-streamlit and the umbrella; graphview is the "consolidated engine". Folding one into the other migrates 4+ dependents and reconciles two panes' goldens. Verdict: DEFERRED — pane merge, wide blast radius; not a byte-safe dedup.

• E — facett-plot ⊂ facett-plotters. ChartKind + Series are duplicated and plotters is a superset — but facett-plot has a real external consumer (facett-warehousedeck) and the two ChartKind enums differ (plot: Line/Bars; plotters: Line/Bar/Scatter/…). Merging is a crate merge that retypes warehousedeck. Verdict: DEFERRED — pane/crate merge with a live consumer; not byte-safe.

• H — tiny format helpers. JUSTIFIED SPECIALIZATION, not forced:

  • human_bytes/human_size (binview/pipeline/plandag/file) share the ÷1024 loop skeleton but each emits a deliberately different string — "1.5 MB" vs IEC "1.5 MiB" vs terse "1.5M" vs a 4-unit cap. One shared fn preserving every contract needs a units+separator parameter at every call site = added surface; one shared fn with a single convention changes displayed text in 3 of 4 crates.
  • crate_color (binview vs coderoom-view) share the FNV-1a hue hash but differ in both the Oklch L/C tuning (0.68/0.14 vs 0.62/0.13) and the return type (Color32 vs coderoom's own Color); merging changes rendered colour (breaks the coderoom hex() robot golden) or needs a parameterised wrapper each site keeps.
  • fmt_dur (cicd, seconds) vs fmt_elapsed (jobview, ms + status-aware) are different signatures/inputs sharing only a ~1-line "{}m {:02}s" tail. Verdict: JUSTIFIED — the skeletons rhyme but the contracts diverge; the workspace already centralises the real primitives (hash_color, look::lerp_color).

• K — gantt vs timeline time-axis. Left at audit time (partial dup; a shared TimeAxis adopter would ADD an abstraction the two visuals don't otherwise share).

Justified splits confirmed NOT dups

Per the audit's "looks-duplicated-but-correct" list: the impl_facet_via_elm! ELM architecture (49 crates, not copy-paste), geomap/osm as well-behaved facett-map consumers, the map3d ↔ map3d-flight layered split (only the terrain DEM, cluster C, was a real copy), the style.rs/WKB/tessellation single owners, and the barnes-hut + graphcloud + cpu-painter core homes.