[HugoXK]

Fixes #2164

Written-by: Dev-2
Owner: Dev-2
Reviewer: Birfy
Claim-Id: a6dd7dd7e74b4328a9bec47d8a8ce344
Branch: dev/2-forward-wire-factory-pitch
Base: ff16097 (main; this head's single parent is 09f8fe7, which was main at push time)
Tip SHA: 3e00ba4
Candidate SHA: b11ae8d61f374132f1a22490f1f2a4a382378c51 (single parent 41ff819)

What was wrong

#2161 (squashed as 96e24ef307e034f3be2706ac60a6631cc8db0e24) made a row-crossing forward wire measure its corridor from the row's lowest measured card bottom. At the factory pitch that bottom can already sit below the next row's top: GAP_Y = 108 (plugins/workflow/src/graph.js:10) and a 120px card standing beside a 48px source puts the row's bottom at 120, four pixels past the target's top edge at 108. The gap clamp then went to 0, the corridor landed on that row bottom, and the run was drawn through the cards of the row below — on one input folding back over itself:

a>c: M 0 48 L 0 120 L -96 120 L -96 108 L 0 108 L 0 108

The same measured row bottom also made #2161's turn depend on a threshold rather than on the corridor: a 92px card beside the source (30–38 latin characters in the shipped stylesheet) leaves 16px of room, and any variant that falls back to the source's own height runs through that card's body and out the other side.

Red first

Three tests were added to core/test/workflows-page.test.js and run on the clean, unchanged tree (git status --porcelain empty).

Against main 41ff8199: 42 tests, 40 pass, 2 fail

not ok 40 - a forward edge into the next row keeps the factory row pitch clear (#2164)
  error: a>c keeps the factory-pitch route (M 0 48 L 0 120 L -96 120 L -96 108 L 0 108 L 0 108)
not ok 42 - a forward edge into the next row turns above the tallest CJK named card (#2164)
  error: a>d turns one pixel above the next row (M 0 48 L 0 112 L 82 112 L 82 108 L 188 108 L 188 108)

The first is the reported shape (the run at the row bottom, inside the cards below, folding back); the second is the reachable 112px CJK card whose bottom reaches past the next row's top.

Against the earlier candidate a0418044 (belowRow = sameRow || room >= 2 * WIRE_GAP, 28px threshold) with the same tests: 42 tests, 39 pass, 3 fail — the same two plus

not ok 41 - a forward edge into the next row turns in the corridor under a taller neighbour (#2164)
  error: a>d runs under the 92px neighbour (M 0 48 L 0 62 L 82 62 L 82 94 L 188 94 L 188 108)

That third failure is why the threshold version was dropped: at 92px (a reachable latin height) it routes at the row top and through the neighbour's body.

At this tip: 42 tests, 42 pass, 0 fail.

What changed

core/src/client/workflows.js (the forward branch of wireRoutes). The corridor is still measured from the row's lowest bottom, but the turn is now decided in three cases instead of by a threshold:

const room = y2 - rowBottom
const gap = sameRow ? WIRE_GAP : Math.min(WIRE_GAP, Math.max(0, room / 2))
const corridor = clearOf(rowBottom + gap, 1)
const cap = sameRow ? Infinity : (room === 0 ? y2 : y2 - 1)
const yA = sameRow ? corridor : Math.max(y1, Math.min(corridor, cap))
const yB = clearOf(y2 - gap, -1)

• room > 0 — the row leaves a corridor, so the turn halves it (capped at WIRE_GAP), exactly the geometry main draws. This is every height a card can reach today.
• room === 0 — the neighbour's bottom sits exactly on the next row's top. The turn stays on that edge, which is byte-for-byte main's route.
• room < 0 — the neighbour reaches past the next row's top. The turn is clamped one pixel above it (y2 - 1), so the run crosses that neighbour and nothing is drawn into the row below.

core/test/workflows-page.test.js gained the three tests above: the factory pitch with 120/140/160 guard heights, the corridor under a 92px and a 96px neighbour, and a 40-character CJK name at 112px. changelog.d/2164-forward-wire-factory-pitch.fixed.md and docs/subsystems/core-client.md record the change, and core/client.js and patches/ds-bot/* were regenerated with node scripts/ds-bot-patches.mjs generate after (cd core && npm run build).

Reachable heights

The shipped stylesheet, measured with its own CSS and markup in headless chromium (card width is a fixed 168px, transform: scale() does not change offsetHeight, hover/selected/status rows do not change the height): latin 1–13 characters 47px, 14–20 62px, 21–29 77px, 30–38 92px, 39–40 107px (a 40-character pure-latin name does not wrap and stays 46–47px, overflowing horizontally); CJK (a 112px-wide name box, 8 characters a line) 1–8 characters 48px, 9–16 64px, 17–24 80px, 25–32 96px, 33–40 112px. The factory cards themselves sit at 47px; across the measured surface the range is 46–112px.

So the reachable ceiling is 112px — a 40-character Chinese name plus its role line, with two independent readings of 112 (Test) and 111 (Product), i.e. ±1px of font metrics. 108–112 is treated as reachable; 120/140/160 are the 113-and-up guard, not a measured shape. (91 is gone: it was the previous round's latin-only ceiling.)

Where it starts

Measured against main 41ff8199 on the factory fixture, one pixel at a time:

neighbour b	room	this tip d	crossings new/main/base
46–106 (61 heights)	62 … 2	byte-for-byte main	0/0/1
107	1	M 0 48 L 0 107 L 94 107 L 94 107.5 L 188 107.5 L 188 108	0/0/1
108	0	M 0 48 L 0 108 L 94 108 L 94 108 L 188 108 L 188 108	0/0/1
109–112	−1 … −4	M 0 48 L 0 107 L 82 107 L 82 108 L 188 108 L 188 108	1/2/1
120	−12	same as 109–112	1/2/1

• b ≤ 106 (every height up to and including the 25–32 CJK band at 96px): the same d main draws, 61 consecutive heights, no drift.
• b = 107 (room = 1) is the first difference, 0.5px: yA is clamped to y2 − 1 (107) where main turns at 107.5. Neither run enters the neighbour, both cross 0 cards, and the 0.5px sits outside the target's border. That height is the 39–40 character latin card, so it counts as reachable.
• b = 108 (room = 0) keeps main's turn exactly — yA = to.top, the same d main draws, 0 crossings on both sides. One pixel higher and the wire would cross a card main does not, which is why this one case is not clamped.
• b = 109–112 is where this fix is better than main, not merely equal: main runs at yA = 109|110|111|112, one to four pixels inside the next row, and crosses two cards (the neighbour above it and the card below); this tip clamps to y2 − 1 = 107, crossings 1 vs main's 2, and nothing is drawn into the next row. At b ≥ 113 the count matches 0bfaf525 (1 = 1) — those heights have no measurement and exist only as the scan's upper-bound guard.

Regression guard

#2161 was a T1 change that reached main with the factory row pitch broken, and the same class of change will come again: any edit to how a row's bottom or a card's prototype height is measured moves this geometry. The three tests here pin the factory pitch (108px rows, 188px columns), a taller neighbour that still leaves a corridor (92px and 96px, both reachable), the exact boundary where the corridor closes (108px, byte-for-byte main), and the tallest measured card (a 40-character CJK name at 112px), with 120/140/160 kept only as an upper-bound scan.

Verification

Ran on the clean, committed candidate b11ae8d6 (git status --porcelain empty), whose eight blobs are the ones in this head (see below):

TZ=UTC ARCH_DOCS_BASE=41ff8199f3c71af442b19b46fded77397f4a78e0 npm run verify

• exit 0. core: 2639 tests, 2639 pass, 0 fail; plugins/social 555 (554 pass, 1 skip), plugins/tree 127 (126 pass, 1 skip), plugins/timeline 140 (139 pass, 1 skip), plugins/workflow 347, .github/scripts 598, interop 37, live-e2e 24 — 0 failures
• arch-docs: OK (8 changed file(s)); ok ds-bot/ = Fei2411/DeepSeek-Bot@2b10c70; ok 10 patches apply cleanly in order; ok core/ equals ds-bot/ + patches/ds-bot/; ok patches/ds-bot/ is what generate writes (generate is a no-op) — build-client --check and ds-bot-patches check are steps inside verify
• generate was run twice, the second run a no-op, and the tree stayed clean; 8 files, +251/−10
• sha256 core/client.js 4ef0b58ae51e8b5dfce243e906a582adabfcfe3309f3fd52c3515b62ac36b88e, sha256 patches/ds-bot/0008-client-bundle.patch 2116890ff54e102ff799312e8e000f090261dfa7fe763b6767e604ae63d8678a — both re-hashed from the blobs in this head, identical to the candidate's

The push rebuilt the commit on top of the main of that moment (09f8fe7e): same eight blobs (all eight blob shas match b11ae8d6), so the diff against main is exactly these eight files with the same +251/−10; the candidate's single parent is 41ff8199, this head's single parent is 09f8fe7e. main advanced once after the push (ff160976, #2166/#2168, which touch only core/src/host/journal.js and its tests/docs), so this head was left as pushed, not rebased: it is still one commit, mergeable, and the diff is unchanged. Not run: nothing beyond npm run verify — no live instance, no browser probe of the served client beyond the test suite.

Checkpoint

{
  "pr": 2172,
  "claim": "Dev-2",
  "issue": 2164,
  "summary": "A row-crossing forward wire no longer takes the row's lowest bottom as its corridor when a taller neighbour reaches past the next row's top. The turn is now decided in three cases: a real corridor is halved as before, a zero corridor keeps main's exact turn, and a negative one is clamped one pixel above the next row so nothing is drawn into the cards below. Tests pin the factory pitch, the 92px and 96px corridors, the 108px boundary and the 112px CJK card against the exact `d` main draws.",
  "files": [
    "core/src/client/workflows.js",
    "core/test/workflows-page.test.js",
    "core/client.js",
    "patches/ds-bot/0008-client-bundle.patch",
    "patches/ds-bot/files/src/client/workflows.js",
    "patches/ds-bot/files/test/workflows-page.test.js",
    "changelog.d/2164-forward-wire-factory-pitch.fixed.md",
    "docs/subsystems/core-client.md"
  ],
  "checks": { "ran": "TZ=UTC ARCH_DOCS_BASE=41ff8199 npm run verify (twice-built, clean tree)", "result": "green", "at": "3e00ba405510a781de662b45be788ebd4464ead0" },
  "next": "none — ready for review"
}

[HugoXK]

@Birfy please review #2172 at 3e00ba4