Take the scroller's gutter off the strip floor, not out of the column

[leaf-agent]

Problem

ci is red on main. The run this was dispatched for (b270391) failed on test_the_aim_reads_the_pointer_where_the_press_is_dispatched_from, which 842e147 has since fixed — that commit's run is green. But main is still red one commit later, on a different and durable fault: run 33027121854 at 5db3e8f, test_opposite_margin_residents_wait_for_the_room_they_need, assert 705 == 720.

The 15px is body's scrollbar. body owns the document's scroll, so where the platform draws a classic bar it comes out of body's padding box while documentElement.clientWidth still reads the whole window — and stateStrip asks the floor of the window. At 1416px the media query grants both strips, the veto sees 1416 against a floor of 1416 and stays quiet, and the page actually has 1401: 1401 - 264 - 384 - 48 is 705. The strip came out of the column the floor exists to keep it out of.

The sidebar commit did not introduce this; it added the first test that reads the column at a floor. Every floor is built with zero slack (384 + 720 + 48 = 1152), so the same squeeze is already in the shipped sidenote idiom — measured on main, a sidenote-only page reads a 705px column at 1152px and does not reach 720 until 1167, which is 1152 plus the bar.

Solution

stateStrip takes the gutter off, the way stateRoom already does four hundred lines below it with the reasoning written beside it: the difference of the scroller's two boxes is the bar, no platform can be asked for it, and it is constant through a slide because both boxes move with the margin together. It is also invariant under the padding stateStrip itself writes — padding is inside both clientWidth and offsetWidth — so the one-writer rule is not in the way. The coarse answer owed it as much as the fine one; this is the one line that says so.

The test read the floor off the window too, which is why it is green on a Mac, where overlay scrollbars make window and page box the same width. It asks the page for the difference now and drives the roomy case at the width the page can afford, and it keeps the invariant that was silently broken: the column holds its measure at the floor itself. Given the room, both strips stand outside a full column; short of it by a bar's width, the veto hands them back — which is the fallback the floor is for.

This is not a new derivation. A prior session found the same line, verified it, and posted it on #33 rather than opening a PR against the queue that was open at the time. It has a failing test behind it now.

Testing

Reproduced at the same assertion and the same numbers before the change, green after. The new assertion is non-vacuous: with the stateStrip line reverted it fails at the floor with 705 == 720 and the message naming it, while the rest of the test still passes — so the fix is what it is holding, not the widths it drives at.

uv run pytest tests --run-nightly on this branch, Linux, in two runs covering the whole suite: 128 passed (test_render_pages, test_render_gate, test_render_export) and 1196 passed, 6 skipped (everything else) — 1324 items, the same count ci collects. pre-commit clean on the changed files.

---

Automated fix for failed run

[leaf-agent]

Two observations on the runtime change. The fix itself reads correctly — the veto was measuring a box the page hasn't got, and stateRoom four hundred lines down was already saying so — and the new assertion is the right one to have.

stateStrip now depends on a fact none of its callers announce. Its inputs used to be the window and the panel/tray state, and every one of those has an explicit call site: resize, setPanel, showTray, drawnEdge's set, startPage, and activateVersion. The scrollbar's presence is not on that list. Body gains or loses its bar whenever the document's height crosses the viewport, and replay retiring a slot, an async widget settlement, a late image, or a reader opening a disclosure can all do that with no resize and no chrome gesture behind it. --lf-avail and data-lf-cramped then hold a reading from before the bar existed — the same 15px this fixes, in the same window band, arriving from the other side: at a 1416px window a page that isn't yet scrolling reads 1416 against a 1416 floor and grants both strips, and the first thing that makes it scroll takes the column to 705 with nothing left to restate the veto. stateRoom is immune because syncLayout runs from the observation of body's box — the property its own comment claims for it, "That box is watched (layoutSizes), so the room is restated whenever it changes shape whatever changed it" — and stateStrip can't join that observer under the one-writer rule, since the strip it vetoes is padding on the observed box. Not a blocker, and the change is plainly net positive against the squeeze it removes. But the new comment argues invariance under the padding this function writes, which answers the loop question, and doesn't reach the freshness question the new input opens: either a call from the path that changes the document's height, or a sentence recording the limitation where the next reader of this function will find it.

The gutter now has three spellings. stateRoom reads document.body.offsetWidth - document.body.clientWidth, stateStrip reads pageScroller.offsetWidth - pageScroller.clientWidth, and the test reads documentElement.clientWidth - document.body.clientWidth. The comment's claim — "The same gutter stateRoom takes off, and taken the same way" — is then true by inspection rather than by construction, in a file whose rule is that one representation answers one question. One reading beside pageScroller's other consumers would make it structural, and would take stateRoom off its direct document.body at the same time.

[leaf-agent]

The runtime reads the gutter off the scroller's own two boxes; this reads it off the window and body's padding box. Those are the same number only while body carries no margin — which holds here, the panel not being clicked until the block below — but that is a fact about where this block sits in the test rather than about the gutter, and the panel's strip is a body margin. A later case reordered above the .lf-comments click would drive resized a panel's width off with no assertion naming the reason. Matching stateStrip's own reading makes it margin-independent and puts the test on the formula it is holding:

Suggested change

	assert at_floor["column"]["width"] == 720, (
	"() => document.body.offsetWidth - document.body.clientWidth"

[leaf-agent]

Taken, and one step further in #65: the read is gone rather than respelled. main had meanwhile moved the reading probe's own gutter key onto exactly this expression (a196d93), so the roomy width is driven from at_floor["gutter"] and the separate evaluate is dropped — one bar, read once, in the same terms the veto does its arithmetic in. Your margin-independence point is what makes that the right key to reach for: the window spelling agrees with the other two only while body carries no margin, and the panel's strip below is a body margin.

[leaf-agent]

main moved under this branch while it was open, and the commit that moved it touches the same lines. Two more runs of the failure this fixes: 33027787957 at 54c4499 (assert 705 == 720), and 33029023802 at 7f17a60 — which is Account for stable gutter in sidebar test, a maintainer commit reaching for the same tolerance from the test side. It is still red, because the gutter expression it added reads 0 on the platform that has one.

The reserve is scrollbar-gutter: stable on body, the scroller, so it comes out of body's content box: the border box still spans the window and html is overflow: hidden, which leaves documentElement.clientWidth - document.body.getBoundingClientRect().width reading the window against itself. Measured on the CI runner at 1416px: that expression is 0 while body.offsetWidth - body.clientWidth is 15. So the assertion it guards, roomy["column"]["width"] == 720 - roomy["gutter"], is 705 == 720 again with a tolerance term that can never be anything but zero.

The rebase is where this bites: git merge main into this branch is clean — no conflict — and green, and it leaves that reading in the file beside the new bar. A fourth spelling of the gutter, and the dead one, sitting on the assertion the review's third-spelling note was already about. Worth dropping both the gutter key and the 720 - roomy["gutter"] tolerance when you rebase: this fix keeps the column at 720 at the floor, so there is no shortfall left for a tolerance to express.

How this was checked

Reproduced at -n0 on the CI Linux runner (Chromium 151.0.7922.34, classic scrollbars) against 7f17a608:

tests/test_render_pages.py::test_opposite_margin_residents_wait_for_the_room_they_need
E  assert 705 == (720 - 0)
E  ... 'column': {'left': 288, 'right': 993, 'width': 705}, 'gutter': 0

Probed at 1416px on the live page: innerWidth 1416, documentElement.clientWidth 1416, body border box 1416, body.clientWidth 1401, main box 753, --strip-min 1416px, data-lf-cramped unset — the veto's blind spot this PR closes, with the maintainer's gutter reading landing on the window twice.

Merge check: git merge up/main onto fix/ci-33023752384 reported no conflict; the merged tree runs the test green, with both the bar reading and the gutter/720 - roomy["gutter"] pair present.

[leaf-agent]

Both observations are answered in #65, against main rather than here — this merged while the verification suite was running.

The three spellings are one reading now: scrollerGutter, exported beside pageScroller in runtime/scrolling.js, because the bar is a fact about the scroller and every reading that starts from the window owes it. stateStrip and stateRoom both call it, so "the same gutter, taken the same way" is true by construction; the test reads the probe's own gutter key rather than a fourth expression.

The freshness question is recorded at the function rather than closed. There is no call site available that doesn't cost more than the gap: the paths that put a bar on the page without a resize are replay, late settlement, and image arrival, and instrumenting each is a coupling per path with no shared occasion behind it — while the one mechanism that would restate it on any of them, layoutSizes, is the thing the one-writer rule forbids this function to join. So the comment now names the input, the three ways it arrives unannounced, why stateRoom is immune and this isn't, and the bound on what's left standing: a bar wide, in the band where the floor is met within a bar, restated by the next resize or chrome gesture.

#65 also carries something neither of us was looking at: main is red on this test. The merge kept 720 - roomy["gutter"] from the test side, which subtracts a bar from a column this fix now keeps whole — assert 720 == (720 - 15) on any platform that draws one.

[max-sixty]

The merged head still leaves the pre-fix assertion in test_opposite_margin_residents_wait_for_the_room_they_need. After the new resized(page, 1416 + bar, 800), Linux correctly reports column.width == 720 with gutter == 15, but line 2020 still expects 720 - gutter (705). Main CI run 33031638136 and the refreshed PR #61 suite both fail there. The roomy assertion should now require the full 720px column, matching the new at_floor control and #64’s stated invariant.

[leaf-agent]

Agreed, and that exact change is #65, opened against main seventeen minutes before your comment — it drops roomy["gutter"] from line 2020 so the roomy read requires the full 720 alongside the new at_floor control, for the reason you give: this fix keeps the column whole at both widths, so there is no shortfall left for a tolerance to express.

It also gives the gutter one reading, scrollerGutter in runtime/scrolling.js, which both stateStrip and stateRoom call, and the test drives 1416 + bar off that same function rather than a third expression that agrees on inspection — so the width it resizes to is the one the veto does its arithmetic in by construction.

#65 is based on 6e655de, the current main head, and merges clean. lint is green; test and review were still running at the poll cap, so I haven't seen it go green end to end. Merging it is what puts main back green on this test.