[ocots]

Part of #314, phase C of the plan. Stacked on #419 (phase B), itself stacked on #417 (phase A).

Conventions: ct-conventions v0.11.0 (SITE rules), ct-agents v0.4.0.

What

Each modeling tab now has a Solution section, and Compare a Solutions side by side section. The solutions are those the convergence crons publish (phase B), and the browser draws them. Julia computes, it does not draw: no plotting library, and nothing is solved at build time.

ProblemPlot.vue, written in plain SVG:

• solves to overlay: the legend chips toggle them. Each chip shows the objective, and an "older version" tag when the solve is stale;
• three views:
  • against time: one small panel per component, in a responsive grid;
  • phase plane: any two components, or time, on the axes;
  • 3D: any three, rotated by dragging or with the arrow keys;
• components: state, control and costate, chosen one by one or by group; "All" shows every figure. By default the state and the control are shown, not the costate;
• a time cursor shared by every view: set by hovering a panel, by the slider, or played (▶, 6 s from start to end). Draw up to t draws the trajectories up to the cursor only, so playing it animates the trajectory as it is traced;
• values at the cursor: in the header of each panel, and as a table (accessibility);
• variable values (for instance tf) listed under the plots;
• colors: the reference categorical palette, validated in light and dark. Each solve keeps its slot whatever is shown, and colors are never cycled. In Compare, the line dash tells the modeling apart, as a secondary encoding;
• in Compare, when a costate is drawn with JuMP among the modelings, a note says that the JuMP costate has the opposite sign ([Dev] Align the costate and dual signs of the JuMP models with OptimalControl #326).

docs/results.jl:

• reads the results branch: git fetch --depth=1 origin results, then git archive read with the Tar stdlib. OCP_RESULTS_DIR=<dir> reads a local tree instead; that is how I built it here;
• freshness, per modeling: up to date when every result carries the current fingerprint of the problem (OptimalControlProblems._fingerprint, the same function the tests use, from test: record the trajectories, the environment and the fingerprint of the solves (#314, phase A) #417), stale otherwise, or no results yet. The badge names the run, its date and its commit;
• Reproducibility block (<details>): the run, the commit, and for each platform Julia, the system, the CPU and the threads. It has a table of package versions per platform, links to download Project.toml, Manifest.toml, versioninfo and Pkg.status, and the command to solve again. It says honestly that the versions resolved later may differ, and that the manifest lists those of the run;
• copies what the pages draw into the site (docs/src/public/results/, git-ignored): one JSON per problem, and the trajectory of one platform per configuration (Linux first). Platforms agree to the solver tolerance, so one is enough, and the site stays small;
• without results (the results branch has no problems/ yet, or no network), the build works and the pages say "No results yet".

ResultLink.vue: a link to a file of public/results/ that follows the base of the deployed version (dev, stable, previews).

Choices

• No chart library. A dependency (Plotly is about 3 MB) would be a second tool, for lines, an orthographic projection and a cursor. The component is about 600 lines and covers exactly what the pages need.
• One platform per configuration is copied, as said above. The trajectories are rounded to 6 digits (test: record the trajectories, the environment and the fingerprint of the solves (#314, phase A) #417), so they barely change from week to week, and gh-pages does not grow every week either.
• The Reproducibility block and the badges are rendered at build time (HTML, searchable). Only the plot is client-side.

Verification

• Data: real solves of test: record the trajectories, the environment and the fingerprint of the solves (#314, phase A) #417, run locally on beam, vanderpol, jackson, robot, glider and moonlander (OptimalControl adnlp+ipopt and exa+madnlp, JuMP ipopt, all converged), published with tools/cron/report.py --results-dir (ci: publish the results of the solves for the documentation (#314, phase B) #419). The documentation was built from that tree with OCP_RESULTS_DIR.
• Screenshots (headless Chrome) checked:
  • each view, in light and dark;
  • a phone width (390 px);
  • Compare with three modelings and the costate note;
  • the Reproducibility block, with its download links answering 200 under the base.
• Build: the results are read in under a second (Results of the crons: 6 problems, read in … s).
• JuliaFormatter (blue) on the Julia files; typos finds nothing.

🤖 Generated with Claude Code