Canvas height now derives from column width via aspect-ratio (CSS custom
prop --canvas-aspect set by JS on the grid host), with --panel-h as a
ceiling. Dropdown options: scaled/locked × 1:1/3:2. Default scaled 3:2.
Legacy 'independent'/'locked' values still parse. Canvas resizes after
aspect changes via requestAnimationFrame.
- /compare accepts ?stem=…&stem=… (repeated) for 2-8 runs; legacy ?a=&b=
still works. compare.js parses multi-stem; template drops stem_a/_b
data attrs that were unused.
- compare-select.js: MAX bumped to 8, button enables at 2-8 selected.
URL emitted as ?stem=… per selection.
- runs list gets a dataset/algorithm chip filter bar above #runs-slot
(pattern ported from metrics.js). Chips reflect the union of values in
the current list; selection state persists across htmx swaps. Non-
matching rows get .filtered-out (display:none).
- _runs.html li now carries data-embedder/data-generator so the filter
can key on them.
- run_args_hash now covers (embed_args, generator_kwargs). When gen_kwargs
is empty we still hash embed_args alone — so plain generators (s_curve,
plain swiss_roll) keep their stems and no existing plain-gen figs need
renaming. Kwargs-bearing variants (swiss_roll_hole, blobs,
gaussian_quantiles, classification) now disambiguate properly.
- Flow persists generator_kwargs into metrics.json meta AND into the
frames.json sidecar meta, so the label-enrichment path can find it
without another lookup.
- _enrich_with_labels discovers gen_kwargs in priority: payload meta -->
sibling metrics.json --> DATASET_META first-match. It matches the
DATASET_META entry by (path, kwargs) so swiss_roll_hole is no longer
confused for plain swiss_roll.
- _cached_frames overrides meta.stem with the URL-requested stem before
enrichment — after a backfill rename the sidecar's baked-in stem is
stale, and we were then failing to find the sibling metrics.json.
- Submit duplicate-check uses the new hash and keeps the hashless-legacy
check as a safety net.
- backfill_hashes.py rewritten: queries Prefect for each recent run's
full params, finds the matching fig under any of (current, legacy,
hashless) names, renames to the current scheme and patches
generator_kwargs into metrics.json.
- panel-grid.js (new): exports mountPanels({host, controls, stems}) → {destroy}.
Moved createPanel + shared control wiring + linked-hover + pad-to-match
time mapping out of compare.js. Stem-count-agnostic; works for 1, 2, or N.
- Panel DOM is cloned from <template id=compare-panel-tpl> on each page.
- compare.js is now a ~10-line shim: parse ?a=&b=, call mountPanels.
- Per-panel color is viridis-sampled by index/N (middle viridis for N=1,
ends-of-palette for N=2, linear lerp for N≤8, cycle at N≥9). Set as
--panel-color on the panel element; CSS reads it for tag/time-seg.
- Homepage <dialog id=run-modal> + run-modal.js hijack the 'embedding' link
(plain click → modal; meta/ctrl/middle-click still opens plotly HTML).
Dialog close disposes every panel's renderer/geometry/material.
- .compare-grid → repeat(auto-fit, minmax(360px, 1fr)) handles N=1..many,
replaces the <900px one-column media rule.
- Runs list: relabel Prefect's 'Late' state as 'Queued' — more honest
description of what the runner is doing at the concurrency cap.
Stem grows an 8-hex sha1 digest of the (keys-sorted) embed_args dict, so
runs differing only in embed_args (e.g. UMAP n_neighbors=5 vs 15) now
produce distinct figs. The stem regex and parser both accept an optional
_<hash> tail so pre-hash figs still render in the runs list and compare
page; legacy filename is resolved on disk fallback.
Duplicate-submission check now rejects against BOTH the hashed and the
legacy hashless variant so users can't accidentally duplicate an old run
either.
Flow additionally writes a <stem>.frames.json sidecar next to the plotly
HTML (same shape as app/web/plotly_parse returns). Server prefers the
sidecar when present; falls back to parsing HTML for older runs. Sidecar
emission is non-critical — any failure just logs and keeps going.
Before dispatching to Prefect, check if figs/<stem>.html already exists.
If so, return 409 with a message asking to bump seed/jitter/N/T or delete
the fig. Stem is derived from (generator, embedder, N, T, J, seed), so
embed_args-only changes still collide and are blocked — users must vary
a stem-shaping param to get a distinct output file.
When two runs share identical params they write to the same figs/<stem>.html,
and the most recent overwrites the earlier. Previously both got a compare
checkbox with the same data-stem, so toggling one toggled both via the JS
Set. Now we flag older duplicates server-side (first occurrence wins — Prefect
returns runs START_TIME_DESC), drop their checkbox, fade the row, and mark
'overwritten' on the outputs line.
Server enrichment regenerates the dataset deterministically (random_state=0,
matching the flow's _DEFAULT_GENERATOR_KWARGS — the stem's seed drives jitter,
not generation) and attaches per-point labels + label_kind to frames.json.
Client picks the dataset-picker's scheme: continuous ramp for s_curve/swiss_roll,
8-color categorical palette for blobs/gaussian_quantiles/classification. Jitter-
added points (id >= num_points) render black. Rainbow material is opaque with
alpha cutoff so overlapping points don't blend to the ramp midpoint.
Swiss_roll and swiss_roll_hole collide on generator_path; the plain variant
wins for now (kwargs aren't preserved through the flow's metrics.json).
Bumped Cache-Control on the frames endpoint so browsers don't cache stale
pre-enrichment payloads.
- applyU now maps u to a shared global frame index uGlobal in [0, maxT-1];
each panel clamps to its own (T-1), so a shorter timeline pads its last
frame while the longer one finishes — both advance at the same wall-clock
tempo instead of rescaling their timelines.
- tick() keeps u as a float closure variable; reading it back from the
integer-step scrubber was quantizing du to 0 at slow tempo + high T
(1600ms/frame, T=24: du ≈ 4e-4 → round to 0 on scrub), stalling playback
after one frame.
Default is smooth (lerp each point between adjacent frames); a motion:step
toggle preserves the snap-to-frame behaviour. Points missing in either
adjacent frame are hidden during the transition to avoid pop-in.
- /compare page renders two ortho-camera panels fed by /api/runs/{stem}/frames.json
- shared controls: play/pause, scrubber, speed (0.5-4x), axes sync (independent/locked)
- linked hover: picks nearest point in one panel, highlights matching point_id in other
- add/remove-jitter nulls handled via per-frame packed positions + setDrawRange
- independent error states per panel; theme-aware colors via themechange event
- per-run checkbox when embedding HTML exists; cap at 2 selected
- sticky 'compare selected' button opens /compare?a=&b= in new tab
- selection state persists across the 3s htmx poll via a Set keyed by stem
- /compare stub validates stems, renders scaffolding (three.js UI next)
Three short paragraphs framing what the notebook studies: stability of
2-D embeddings under controlled perturbation of the input over time,
the two metrics logged per run, and why the streaming/longitudinal
angle matters for both visualization and downstream classification.
Read --picker-panel from the card's computed style at scene creation
and re-apply on themechange, so the three-js canvases flip with the
rest of the UI instead of staying stuck on the light palette.
Client-side: disable the form submit button until the picker writes a
dataset_id. Server-side: reject posts where dataset_id is present but
empty, instead of silently defaulting to make_s_curve.
