Flows

The board is the template, drawn — not a picture of it

The brief said a flow already is nodes and edges, so the board should render the data model honestly. I took that literally: board.js reads the stored flow document in the spec’s §3 shape and lays it out. Steps become boxes labelled with their seat, order becomes arrows, the gate field becomes a diamond, the loop edge becomes a back-arrow. Nothing is hand-positioned, so nothing can fall out of step with the template.

Which also settles the read-only question structurally: there is no authoring surface because there is nothing to author into — the picture is downstream of the data.

Pick by shape only works if the shapes actually differ

“Choose by shape, not by reading a list” is only true if the three v1 flows look meaningfully different at a glance. They do, and not by decoration — by their real structure: Readiness is a single node ending in a report and never writes; Scene Polish is one node into a gate; First Act is a four-node chain with a gate and a loop back for blocking annotations.

If a future flow duplicated another’s shape, picking by shape would quietly stop working. Worth knowing before the flow set grows — flagged below.

One state, computed twice — not two states kept in step

The hard guardrail is that the board and the rail must never be two systems. I did not implement that as discipline. There is one run object, the board draws it, and the rail is derived from it by deriveRail(), which reads the current step out of the template and returns the pace, seat, loop state and step count.

The rail cannot contradict the board because the rail does not store a pace — it asks for one. Both columns on the comparison page are driven by that single object, and a test steps the run through every position and asserts the two agree at each one.

“Replace” is the wrong verb, and that is the answer

The board and the rail are not two views of one thing — they are two zoom levels of one thing. The board says which step of the route; the rail says what is happening inside that step. Neither is redundant, so replacing one with the other removes information rather than removing a duplicate.

The honest version of “replaces” is therefore nesting: the rail moves inside the open node, which is what it always was. I built that as option A and it is the more elegant idea — but at 300px it costs a level of containment on the busiest surface for no information gained. I recommend B, and because both are drawn from the same state, changing later is a layout decision rather than a rewrite.

The two-cast problem does not get solved. It dissolves.

Round 1’s hardest visual problem was making Floor-cast and in-app-cast unmistakable — a triangle and a dot, never colour alone. Residents-only removes the premise: there is one agent model, so there is one mark. .cmark.resident keeps the shape+text rule (a diamond mark, never colour alone — still has to survive greyscale and a screenshot) but the system it belongs to is simpler by construction, not by discipline.

Round 1/2’s .cmark.app / .castbar.floor rules are still in flows.css, unedited — On the rail still renders them. That is honest history, not dead code to clean up.

The front door was the hero twice. This round it is not the door at all.

Keystack-first means most sessions never touch the front door — a resident starts the run, and the human’s first frame in 4P is mid-route. The actual design problem this round is orientation from a cold arrival, not a picker. Arrival answers four questions in one glance — which shape, which step, who is in which seat, what needs you — and the door becomes the page you visit when nothing is running, not the page everyone starts from.

Five shapes need five silhouettes, proven the same way as round 2

DEVELOP/DRAFT/REVISE/TRANSFORM/JUDGE reuse board.js unchanged — a shape is a flow-document template exactly like round 1’s three flows, so the renderer needed one small extension (a second lookup table, tried after FLOWS) rather than a rewrite. The five signatures (nodes|gate|loop) are 2|T|F, 3|T|F, 4|T|T, 4|T|F, 2|F|F — asserted distinct the same way round 2 asserted the original three, because round 2 already flagged that picking-by-shape silently breaks on a collision.

REVISE keeps the round‑1/2 revise‑loop (blocking annotations); TRANSFORM does not loop — that is the one deliberate difference that keeps their otherwise-identical node counts (4 and 4) from colliding.

The proposal card is not illustrated. It is computed against a real roster.

“The Salt Line” has two residents — Noa writes, Wren edits — and no Continuity resident. That is not staged for the demo: it is the actual roster residents.js declares, and resolveCast() checks every card’s required seats against it. DRAFT reads castable. REVISE and TRANSFORM both read blocked, honestly, on the same missing seat — but not identically: REVISE has a declared solo variant (Scene Polish, Editor only) so its fix offers a way to run today; TRANSFORM has none, so its fix is only add a resident. Two blocks that look the same until you read them is exactly the kind of flattening round 1’s readiness report was built to refuse.

Busy is rendered as its own, third fact — not a weaker block. Wren mid‑run elsewhere makes the DRAFT proposal queue, not fail; the same resolver call, one flag different, produces a genuinely different verdict.

The honesty rules force a fourth verdict

The spec names three: PASS / WARN / BLOCK. But rule two — a check over nothing says no data yet, never a confident green — cannot be expressed in three, because a vacuous pass is a pass unless the absence of a measurement has its own identity on the page.

So there are four. NO DATA is drawn colourless and dashed: not a severity, not a failure, an absent measurement. It can never be mistaken for a green at a glance, which is the whole point. The verdict banner also asserts the count first — a “can run” conclusion is only reachable when a positive number of checks actually returned something.

And the banner needs a third state, which I only found by testing it. My first build drew the per-check honesty correctly and then committed the exact sin at the top of the page: for the television case it printed “This flow can run” in the largest type on the surface, because nothing was technically blocking — the gating check had simply never been measured. A vacuous pass, promoted to the headline. The rule that fixes it: NO DATA on a check that would otherwise gate the run is not a silent yes, so the banner says “we cannot tell”. Three banner states, not two.

The cast mark rides on the attribution, not the header

The spec puts the cast in the room header. I have done that — and then also put it on every seat attribution, because a header-only mark fails the moment one note is quoted, exported, screenshotted or pasted into Slack. At that point the header is gone and a Floor-agent line reads as the product’s persona, which is exactly what risk 3 forbids.

The mark is shape and text, never colour alone — a triangle and the words Floor agent, versus a dot and In-app. It survives greyscale print, colour blindness and a screenshot.

Both doors are the same size

The fold ruling says Flows is the room’s front door, one user-facing feature. The risk in building an objective picker is that it quietly becomes the only way in, and freeform — which the spec says “remains exactly as it is” — degrades into the fallback nobody chooses.

So the two doors are the same size and sit side by side, and the freeform card says in as many words that it is unchanged. Flows is offered first by position, not by making the alternative look like a downgrade.

The report opens onto the stage, not into a 300px dock

A readiness report with per-check fixes and a coverage chart does not fit the dock, and squeezing it in would produce exactly the readiness theater the spec warns about — a report too cramped to read is a report nobody reads.

So the dock summarises and the stage expands, which is the pattern the gate tray and continuity lane already use in v2. No new surface, no new navigation: the same Open gesture that already ships.