Experiment: static vs dynamic parts, and the Slot that lets the caller choose

[erykciepiela]

Scan of every MDC member recording which of its parts are static (fixed at construction), dynamic (read from the channel), or want to be both — plus a prototype that lets the caller decide. Full note: doc/experiment-static-dynamic-parts.md.

The scan

Genuinely static: typography, elevations, cardActions, chipSet, list, listItem, dataRow, dataCell, layoutGrid, divider.

Genuinely dynamic: the three progress displays, snackbar/banner messages — snackbar is the only component whose text part is dynamic by construction.

Want both, only static available — 16 members:

Part	Members
label / caption / title	toggleSwitch, filterChip, button, fab, menuItem, menu, card, topAppBar, dialog, simpleDialog, drawer, tooltip
icon	iconButton, button, fab
option list	select, radioButton, segmentedButton, tabBar
geometry	slider min/max, dataTable columns, imageList columns, layoutCell span
content	checkbox's label widget (pinned at {}), imageListItem src/label, text-field helper/error text (no path at all)

Two of these the codebase already pays for: iconToggle is an entire component that exists because iconButton's icon can't follow data; and photo-gallery couldn't build an image list from data, so it used displayed $ dynamic \m -> each …, tearing down and rebuilding the whole list on every album change — two static parts cost a reconciled collection.

The API

The staging is already there — construction (config args) → registration (body, chrome on screen) → feed (with/mvu); body $ w is already the no-data peek and body $ with sample $ w the sample-data peek. What was missing is per-part choice of stage:

data Slot i a = Absent | Pinned a | Fed (i -> a)

card { caption: "Quiz" }        -- Pinned: renders at registration, no channel
card { caption: fed _.album }   -- Fed: renders on first feed
card {}                         -- Absent: no caption node

Pinned a is deliberately not Fed (const a) — only Pinned is visible without data, so the type records what is peekable.

Measured

card and topAppBar converted:

• All 31 demos compile unchanged, no new warnings — the lifting rides the existing ConvertOption tag mechanism.
• A widget stays polarity-agnostic until a call site actually asks for a dynamic part; the cost is that card/topAppBar are no longer Ocular (precedent: drawer, attrWith).
• Browser-verified in a new scripts/smoke/tests/photo-gallery.mjs: the fed title updates its own text node in place — a slot is a channel, not a rebuild.
• spago test and the full smoke suite pass.
• photo-gallery's app bar now names the open album, retiring its separate headline2 … # tapped stage.

Findings

1. The dynamic case must be named. { caption: _.album } does not typecheck — an instance chain can only commit on a concrete from, and _.album is still { album :: t | r } -> t at conversion time. Hence fed. Arguably a feature: the config record then says out loud which parts wait for data.
2. Absent can't fold into the slot — whether a part's node exists is construction-time knowledge, unavailable to a function of i; the three-way split preserves today's behaviour exactly.
3. Today's static/dynamic boundary coincides with the mvu seam. Every demo writes card {…} $ (pipeline # mvu seed), so chrome is structurally never fed. A dynamic caption forces (card {…} $ pipeline) # mvu seed — the one call-site restructure, visible in the photo-gallery diff.

Alternatives considered and why they lost (twin words, widget slots, row promotion), plus follow-ups in value order, are in the note.

🤖 Generated with Claude Code