The block editor: the only stateful module in the package's rendered half (ADR-0005 decisions 6, 8, 9, 13).

Everything this component does is translation. Every author gesture reduces to one of the four commands StatifierBlocks.Edit defines, and the mutation, its inverse and the set of places a block may be dropped are all pure functions of {document, palette} that were tested with LiveView absent from the dependency tree. What is left here is turning a phx- event into a command, offering it to StatifierBlocks.Edit.History, and re-rendering.

That split is the single most load-bearing constraint in ADR-0005, and the reason is drag-and-drop specifically: it is the interaction most likely to be tested by clicking around and declared fine, and the one most likely to corrupt a document when it is wrong. If the semantics of a drag lived here, they would be testable only through a browser driver.

The drag, in two round-trips

dragstart pushes one event. Edit.Targets.slot_verdicts/3 runs once, the result goes into the drag session, and the re-render stamps data-drop on every slot - so every valid target highlights before the pointer has moved, and hover costs nothing. drop pushes the position, one :move is built and applied, and the tree re-renders. dragend clears the session. There is no third round-trip and no client-side validity logic.

That one enumeration answers both halves of the drag. The accepting slots are data-drop; the refusing ones that have a data-flow reason to give (the 2026-08-29 ADR-0003 amendment's vocabulary) carry it as data-drop-reason beside it, so a later hover affordance has the reason already in the markup and still needs no round-trip and no JavaScript. The host's widening relation reaches all of this the way ADR-0003 decision 6 says it does and the way nothing else could: through palette.assignability, consulted by the one Assignability the compiler's validate/3 consults.

The other thing a drag can carry (sb-4nep)

A palette entry is a drag source too, and dragging one onto a gap inserts a block of that type there. It is the same two round-trips with the same one enumeration: insert-dragstart builds the session against a probe block of the dragged type instead of against a block the document holds, and insert-drop hands the gap's position to the same insert_from_palette/3 a pick uses. The drop therefore produces decision 2's :insert at a position the author named, exactly as the "+" and the pick do, and ADR-0005's command set is untouched - what is new is a gesture, not a command.

Which is also why the position comes from the gap rather than from palette_position. Arming is how the click path names a destination; a drag names one by landing on it. A drop made while some other gap is armed lands where it was dropped and clears the mode, because the gesture that just finished is the one that said where.

Config, and the gate that keeps the document sound

ADR-0002 decision 6 guarantees slots/1 returns without raising only for config validate_config/1 accepts - and an author halfway through typing an identifier has invalid config almost continuously. So an :update_config reaches the document only when it validates, which Edit.check_config/3 enforces inside History.commit/4.

In-progress form state that does not validate lives in this component's drafts assign: never in the document, never on the undo stack. It is overlaid onto the selected block's form at render time - values and findings both - so the author keeps their keystrokes and sees findings about the value they are currently typing, while the tree, the slot set and every consumer downstream still see the last config that validated. Crucially, the overlay never calls slots/1: it touches the form and nothing else, which is what keeps the draft from reaching a callback that was only promised valid config.

The datamodel, and why it is one assign and no logic

A host may hand the editor the datamodel paths it declares. The only thing that buys is the undeclared-path advisory ADR-0005 amendment 11e-11g specifies: a config field a block type annotated datamodel_path?: true whose value is not in that set gets an :info finding anchored on the field. The rule itself is StatifierBlocks.Datamodel's and is pure, so it is tested with LiveView absent; what happens here is one normalization on update/3 and one concatenation in rebuild/1, which is the same translation-only posture as everything else in this module.

nil is the default and means no datamodel supplied, which per 11f produces nothing anywhere - the check does not run at all. That is not the same as an empty set, which is a host declaring that its documents address nothing.

A datamodel is no longer the only thing that declares. ADR-0005 amendment 11k reads three sources, and two of them are roots rather than paths: the roots the document itself declares (ADR-0001 decision 11), which StatifierBlocks.Datamodel reads off the document with no help from here, and the roots the compile call names. The second is why there is a declare assign: this component has no compile call to read, so a host that will pass declare: to StatifierBlocks.Compiler passes the same list here. It is the same one normalization and one concatenation the datamodel gets - [] is the default and declares nothing, so a host that never passes it sees exactly what it saw before.

The run marks a host paints

A host replaying or executing a document has two different things to say about it on the canvas: where the run is - the blocks a step has activated - and who it is waiting on - the block whose call out to a handler has not come back yet, and how it came back once it did. Those are active_marks and invoke_mark, and they are two assigns rather than one because a step can carry either without the other. One seam that took both would make every caller holding one of them pass a placeholder for the other.

They are assigns and not a new API, because every other thing a host tells this editor is an assign. A host that re-renders for its own reasons can pass them in the component call; a host reacting to a run event it received out of band pushes them, which is send_update/3 and needs nothing from this package:

Phoenix.LiveView.send_update(StatifierBlocks.Editor,
  id: "editor",
  active_marks: ["blk_capture"],
  invoke_mark: {"blk_authorize", "done"}
)

Which is exactly why a mark is held as editor state rather than read out of the assigns at render time. send_update/3 delivers only the keys it names, so a component whose marks lived only in the assigns the host most recently passed would drop them on the next unrelated re-render - an author moving a block, a header the host redrew for its own reasons. Each mark is written only by an update that carries it, the way datamodel and drawer_height already are, so a render that says nothing about the marks changes nothing about them.

A different document clears them, and that is the one place they part company with the pane folds. ADR-0005's 2026-08-30 amendment to decision 2 exempts a fold from the document-switch reset because a fold addresses no block; a mark addresses exactly one, so it stops being true the moment the block it names is gone. Marks reset in switch_document/2 beside selected_id and the collapsed set. A host that swaps a document and marks the new one in the same update is not fighting that reset: the reset runs first and the marks it passed are applied after.

The active marks are also what the toolbar's Fit active falls back on. An observer watching a run is not selecting anything - the marks are the run's own answer to "which block matters right now" - so the control is enabled whenever there is a selection or a mark, and with nothing selected it fits and reveals the first marked block. The order is the host's: the marks are kept as the list the host passed, deduplicated only where the canvas asks the question as a set. With no selection and no marks there is nothing to fit, and the control stays disabled.

What reaches the markup is data-run-active, data-run-invoking and - only for a call that has come back - data-invoke-outcome, on the block's .sb-node. The outcome is passed through rather than checked against a set: which outcomes a call can have is the block type's vocabulary, the one slot_outcome_key already reads, so a closed set here would be this package inventing a second one. The stylesheet tints the two the spike proved and leaves every other outcome the neutral treatment.

Opening at a fit (sb-ehqn)

A document wider than the canvas opens with its right-hand columns off the edge, and the only remedy the editor had was the author pressing Fit width on every document they opened. fit is the host's opt-in to having that press made for them: :manual (the default) is exactly today's behaviour, and :width or :active makes the editor open the way it looks after that button.

It is an opening state and not a control, which is the whole of the care in it. The fit needs numbers only the browser has, so it cannot happen at mount; what happens at mount is that the mode is set and a fit is armed, and the first measurement payload spends it - the same computation handle_event("fit", ...) runs, on the same ladder, against the same measured scroller. It is spent once per open document, and re-armed only when the host swaps a different document in: a host re-renders for reasons of its own, and an attr that re-fitted on each of them would throw an author back to the fit every time their own header changed. What the attr opens is a document, though, so a different document is another opening and is armed from the attr the host passes in that same update - :manual or an absent attr arming nothing, exactly as at mount. Between one document's measurement and the next document's arrival the attr is inert, zoom -/+ return the canvas to :manual as they always have, and the editor is in the state it would have been in had the author pressed the button themselves.

A host that never imports the measurement hook measures nothing, so the fit is never armed away and never spent: the mode is set, the canvas is at 100%, and that is decision 7's absent-hook test holding here too.

Between the mount and that first payload the canvas is laid out at 100%, which is a frame the author should not be shown: it paints the whole chart at full size and then snaps to the fit. So for exactly as long as a fit is armed the root carries data-fit-pending, and the stylesheet keeps the stage unpainted under it - the layout still happens, because the layout is what the hook has to measure; only the ink waits. The stylesheet also carries a delayed reveal so the wait cannot outlive the frame it exists for: a hook-less host never spends the fit, and a blank canvas forever would be a worse defect than the flash. Nothing about this is the hook's - it writes no attribute, no style and no class (decision 7a), and the attr is server-stamped like every other.

What stays the host's

Which palette entries a tenant may use, who may edit or publish a document, where it is stored, and what publishing means are all outside this package (decision 15).

The 2026-08-29 shell amendment's ruling 8A adds two more, and names the seam for each: slots for markup, events for actions. The outer header - document identity, the document switcher, the theme control, compile and publish - is the :header slot, because markup is exactly the thing a host wants to own and a slot costs this package no API surface. The drawer's height is on_drawer_resize out and drawer_height back in, because the height is remembered per viewer and this package has no viewer. A host that renders its own publish button into the slot and receives the press as its own event is the intended shape; the editor draws no header of its own and is not waiting for permission to. So is concurrency: this is a single-session component, it surfaces the revision it loaded so a host can do optimistic concurrency on save, and it does not merge, rebase or resolve anything.

The drawer tabs a host contributes

8A's split gives the drawer to the package, and 1A's test - tabular, and about the whole document - governs what the package puts in it. Neither anticipated a host with content that passes that test: a host executing the open document has a run feed, and a feed of steps is a grid of rows about the whole document. The 2026-08-30 amendment recording the drawer's tab strip as a host seam transfers 1A to the host for the tabs it contributes, and drawer_tabs is where they go.

Each entry is %{id:, title:, content:} with an optional count:. id is the host's own name for the tab and is what its DOM id and its panel's are built from; title and count are what the tab and the collapsed strip draw; content is a function component - the same seam shape icon and expression_component already use, for the same reason ADR-0005 decision 9's does: HEEx has no dynamic-component tag, and a one-argument function returning a rendered struct is the whole of the contract. It is called with the tab's id and count, and with change tracking on those two keys only - which is why the example reaches for assign/2 rather than Map.put/3. StatifierBlocks.Editor.Drawer's moduledoc has what goes wrong when a host merges its own values in instead.

Phoenix.LiveView.send_update(StatifierBlocks.Editor,
  id: "editor",
  drawer_tabs: [
    %{id: "runs", title: "Runs", count: length(events),
      content: fn assigns -> MyApp.run_feed(assign(assigns, :events, events)) end}
  ]
)

A function and not a slot, and the reason is the live feed the seam exists for. A slot's body is a closure over the host's assigns, which sounds like the shorter path to live content and is not: a LiveComponent is re-rendered when the assigns it was passed change, and a host assign read only inside a slot body is not one of them, so the appended step never reaches the screen. Pushing the descriptors is how every other thing a host tells this editor arrives, send_update/3 included, and it is live for the same reason the run marks are.

Held as editor state, like the marks and for the same reason: an update that says nothing about the tabs changes nothing about them. A host tab named for one of the package's own tabs, or repeating an id already used, is not drawn - StatifierBlocks.Shell.host_tabs/1 says why.

The findings number a host may show (sb-ukgu)

A host that draws its own header usually wants to say how many findings the open document has, and the obvious way to get that number - counting whatever list the host itself passed in, or counting the compiler's raw output - produces a different number from the one the drawer's Findings tab reports. It has to: the drawer counts the caller's findings plus the :resolution and :config findings ViewModel derives plus the undeclared-path advisories, and the host's own list is only the first of those three. Two numbers for one document, side by side on the same screen, is the defect this seam closes.

So there is one number, the drawer's, and findings_count/3 is how a host reads it:

StatifierBlocks.Editor.findings_count(document, palette,
  findings: findings,
  datamodel: datamodel
)

Its arguments are deliberately the assigns the host already holds and already passes to the component, not the editor's internal state. That is what makes it usable: the editor is a LiveComponent, so a host has no handle on its socket, and a number that could only be read out of that socket would have to be pushed back through on_change - late by one render on mount, and absent entirely for a document nobody has edited yet. A pure function of the same inputs is available on the host's first render, needs no round-trip, and cannot drift: the option keys are the assign names, and the component's own rebuild/1 builds its view model through the very same private function this one calls.

What comes back is Shell.findings_count/1 over ViewModel.findings - orphans included, for the reason recorded there. A host showing a different number than the drawer is a bug in the host; a host showing none is fine.

The insert mode, and the pick that lands nowhere (sb-dfyk)

palette_position is a mode, and every visible part of it hangs off that one assign: the armed gap on the canvas, the instruction naming where the pick will land, the Cancel beside it, and the Escape binding - which is on the root only while the mode is open, so a resting editor holds no window listener for a mode nobody opened. Arming also un-collapses the palette, because a mode whose only instruction is inside a folded pane is the same defect in a different place.

A pick made with nothing armed stays a no-op, and that is a ruling rather than an omission. The alternative on the table was appending to the selected container's default slot, and "default slot" is a rule no accepted ADR states: decision 8 ties a pick to a position the author named at a gap, and inventing a destination on their behalf would be a new contract written into a handler rather than into the record. What the no-op was missing was not a destination but a reason, so it now gives one - palette_unarmed_pick is that sentence, and it renders in the same region the armed case uses for its instruction.

The selection a host can follow (on_select)

The selection is editor state: it is produced by a gesture on the canvas and only the component knows it, so - unlike the findings number above - there is no pure function of the host's own assigns that answers it. What cannot be read has to be pushed, and on_select is the push.

It is a one-argument function, called with each new selection and never for its return value, and it sits beside on_change rather than inside it: a document and a selection are different subjects, and a host that wants one should not have to receive the other. What arrives is a descriptor rather than a block -

• because the host already holds the document it passed in, so returning the block's config would be a second channel for something the host can already read, and one that goes stale the moment the two disagree.

Deselection calls on_select with nil, because a panel that follows the canvas has to be able to empty itself; a callback that only ever fired on a new block would leave the panel showing the last one forever. It fires when the selection changes and not otherwise - not on every render, not on an edit to the selected block, not on a mount with nothing selected, and not on re-selecting the block that is already selected.

Nothing about it is a document edit. There is no :select command, and nothing here is serialized, stored, undone or redone. The package's own inspector still reads the selection out of component state: this is a seam out, not a rewiring of what is already inside. See ADR-0005's 2026-09-05 amendment, the host seams, on_select and a selection descriptor.

Assigns

| value_candidates | no | the values offered per datamodel path, %{path => [%{label:, value:} | binary]}; merged over the datamodel's own one_of enumerations, per path, so a path this map names uses this map's list and a path it does not name keeps what the datamodel declares. Read only by an expression editor that draws value pickers; %{} (the default) now means nothing beyond what the datamodel declares rather than nothing at all | | invoke_types | no | the invoke types the host is prepared to answer; suggestions on an invoke_type field, never a constraint, and [] (the default) means no list supplied | | chart_outcomes | no | what the host says each of its stored documents finishes with, %{document id => [outcome]}; offered as candidates on a core.subchart's outcomes field and compared against what that block declares (StatifierBlocks.ViewModel.outcome_findings/3). %{} (the default) says nothing about any chart, and a chart the map does not name is unknown, which is not disagreement | | active_marks | no | the block ids a run has activated; held as editor state, and cleared when the host opens a different document | | invoke_mark | no | the block a run is calling out to and how the call came back - {block_id, outcome}, a bare block_id for no answer yet, or nil for no call at all | | theme | no | --sb-* custom properties for the canvas root | | fit | no | the fit the editor opens in: :manual (the default), :width or :active; the first measurement performs it once, and an unknown value is refused into :manual | | fixtures | no | %{block_id => [TruthTable.t()]}, read by both the drawer's truth-table tab and, as of sb-4yze, its Fixtures tab (refresh_fixture_runs/1 drives each row through the compiled chart), and, as of sb-e30x, by the inspector's config form for the fixture hint beside an :expression control, and, as of sb-0l36, by the inspector's own Fixtures tab, which shows the selected block's runs out of the same result; nil (the default) means no fixtures source, and the drawer is still there with a count of 0 | | drawer_tabs | no | tabs the host contributes to the drawer, each %{id:, title:, content:} with an optional count:; drawn beside the package's own and rendered by calling content | | drawer_height | no | the drawer's height in rem, remembered by the host per viewer (2A); bounded on the way in | | on_drawer_resize | no | one-argument function called with each new drawer height, which is how the host comes to have one to remember | | class | no | appended to the root element's own classes | | history_limit | no | bound on the undo stack; :infinity by default |