ui: phase-portrait / orbit lab window (fleet UI ladder, rung 1)

[InauguralPhysicist]

From hq/proposals/fleet-ui-expansion.md (DRAFT, awaiting CEO go on ladder order). Rung 1 — the most visually-natural domain in the portfolio currently opens no window.

Scope: a run() front-end on the existing pure core — live orbit + phase-portrait rendering, parameter sliders, bifurcation sweeps, and observer predicates displayed as live forecasts (the observer IS the instrument — this is the most on-thesis UI in the portfolio).

Forcing contract: chart/plot widget as a THIRD lib/ui consumer (with eigen-sheet#26 and EigenMiniSat#76 — design the widget against all three at once), continuous-redraw canvas, pan/zoom interaction.

Standard from birth (see Tidepool#43 / eigen-edit#3): theme layer (in-place field-by-field apply before widget construction), render-decode + mouse oracles alongside the UI, desktop-gui-qa before/after screenshots as acceptance.

Oracle: headless run vs UI-stepped run of the same system/seed produce byte-identical trajectories — the UI must not perturb the math.

Owner: eigenscript-app-fleet-engineer; upstream gaps to eigenscript-ui-toolkit-engineer. Gate: portfolio coverage-map row on CEO go.

[InauguralPhysicist]

Gate cleared: CEO go 2026-08-02, ladder order as filed — hq/proposals/fleet-ui-expansion.md is APPROVED and the coverage map (hq/PORTFOLIO.md, lib/ui lane) now carries the program. Rungs proceed one at a time; this rung starts when the previous one is green (suite + oracles + desktop-gui-qa screenshots).

[InauguralPhysicist]

Review finding from slice 1 (PR #21) — carry into slice 2, non-blocking.

sim.pts grows without bound in the interactive path. frame_advance appends every frame forever; TRAIL caps only what _paint_phase draws (it's a start index, i0 = n - TRAIL), not what's stored. In auto/dump mode that's correct and load-bearing — traj_text needs the full trajectory for the oracle — but an interactive session (frames < 0, runs until Escape) accumulates ~3,600 two-element lists per minute at 60fps, order 20MB/hour. Leaving the lab window open is a slow, silent climb, which is exactly the shape the constrained-machine discipline exists to catch on a 4GB box.

Suggested fix for slice 2: keep the unbounded history only when dumping (auto_frames >= 0); in interactive mode trim sim.pts to the draw window. Whatever the fix, it must not touch the auto/dump path — the byte-identical oracle depends on the full trajectory, so a naive global cap would break it.

Also verified during review (recording so slice 2 doesn't re-litigate): the oracle is genuinely non-vacuous — the UI half drives a real window, real widgets, and a real per-frame draw through ui.app_loop, sharing only frame_advance/traj_line with the headless half — and it holds byte-identical at zeta = 0.137, 0.3333 and 1.75, not just the harness's hardcoded 0.15, so the build-time _on_zeta_change slider write-back doesn't quantize.

[InauguralPhysicist]

Slice 2 merged (PR #22). Rung 1 status: items 1–3 of the slice-2 scope are done — history bound, pan/zoom as view-state-only, and a real mouse/render-decode oracle.

Verified independently before merge (not taken on report): test_orbit_hist.sh and test_orbit_oracle.sh both re-run locally on the branch — all four oracle legs pass, and both planted faults (zeta perturbation; cap removed → 5001 points retained) are caught. The cap is correctly confined to the interactive branch (frames < 0), so the dump path still keeps every point and the byte-identical oracle is untouched.

Gaps routed upstream rather than left in FINDINGS.md: EigenScript#822 (wheel events carry no cursor position — cursor-anchored zoom needs a pointer shadow and is wrong for a wheel before the first motion event) and EigenScript#823 (canvas doesn't clip its own on_paint; labels overflow their column). #823 is the one that caught the truncated side-panel labels that had been shipping since slice 1 and were invisible to every headless test.

Remaining for rung 1: the bifurcation sweep view (deferred from slice 2 to keep the diff reviewable). That plus the x-y plot widget landing upstream (#819) would close this issue.

[InauguralPhysicist]

The x-y plot widget has landed upstream — EigenScript#819 (PR #824), merged 2026-08-02, all 18 CI checks green including asan+ubsan at a zero leak tally.

chart was generalized in place into a data-coordinate x-y plot (no parallel plot type; a series with no x keeps index-x, so nothing regressed). It was designed against all four waiting consumers at once — this issue among them — so the API should already fit: named multi-series each with its own x list, fixed_aspect, labelled vline/hline/point markers, add_point/add_xy/chart_trim for live incremental data, and drag-pan/wheel-zoom owned by the widget (it reads the toolkit's own pointer state, so no consumer has to shadow mousemove — see #822).

Clipped and theme-keyed by construction: the suite asserts zero primitives outside the widget rect under extreme pan and zoom, planted-fault validated, and confirmed in real pixels (the ring around each chart decodes to exactly the panel background). API reference is in docs/STDLIB.md.

This consumer can drop its hand-rolled canvas plotting whenever the next slice touches that code. It arrives on the next EIGS_REF bump.