Fix margin rail, float placement, thread marks, and preview tab order after the chrome change

[leaf-agent]

Problem

Run 33327310444 on d2dd825 failed with 34 tests red. Most of them are owned: #121 has the scrollbar-gutter set, #132 the generic CONTROL contract, #134 the catalog stills, #136 the sidebar width, #137 the versions menu, and #133 tracks the digest. What was left unowned is five tests that all fail on one surface — what the panel's opening does to the page around it — and each is a reading 335bf9a "Delegate page chrome to browser standards" moved without a second reading moving with it. All five pass at 335bf9a's parent; each was reproduced at -n0 before it was touched.

Solution

Four readings, each corrected where it is taken. No assertion is loosened and no test is changed.

The margin's 900px floor asks the window again. That number is half of a pair: @container lf-shell (min-width: 900px) grants the rail and the marker's strip, and @media screen and (max-width: 899px) stops drawing the living margin at all. 335bf9a converted the grant to a container query and left the hide a media query, so the two came apart the moment the panel narrowed body: at 1200px with the panel open the shell is 765px, the grant lapsed, and the markers stayed on screen with no strip reserved for them. Measured on SUGGESTION_PAGE, --strip-r resolved to max(0px, 0px, 0px, 0px) and every suggestion's control row docked beside an empty rail. The floors below it stay container queries: 1152 and 1416 ask how much room is left beside the panel, which is the question the runtime's old data-lf-cramped veto asked, and their answer is supposed to move with the panel.

Body's inline size is back under the layout observer. 335bf9a dropped layoutSizes.observe(document.body), and with it the only thing that re-ran syncLayout while the panel's eased margin was still narrowing the page. setPanel runs one synchronous pass at the press, reads the wide box, and returns; nothing reads the narrow one. So a composer placed in a wide window's margin kept a place the narrowed page no longer had — an absolute child past body's client box, which is sideways-scrollable overflow, with the box standing on the panel that displaced it. Restored with the content-box width filter the removed code carried, so ordinary document-height growth still does not feed back into a writer that reserves flow content.

The panel layer is shown before anything is rendered into it. The panel is a <dialog> now, and a dialog nobody has shown is display: none — but setPanel called renderPanel() first and syncPanelLayer() (which calls show()) second. renderPanel is where the anchor pass runs for the threads it draws, and a mark hangs on the boxes its element shows through, so a widget an agent sent in a reply resolved to an element measuring 0×0, took no mark, and never got one afterwards: the thread stayed open in the panel pointing at nothing on either side. Swapping the two lines is the whole fix; the focus target is threadsBox, which the render fills rather than replaces.

The margin preview shows without naming the marker as its source. showPopover({source}) puts the card in the sequential focus order directly after the control that invoked it — but this card is not invoked. It opens on the marker merely taking focus, so a reader walking the margin was made to Tab through a close button and two preview actions they had not asked for, three stops between one suggestion's controls and the next, with the card's contents changing under them as they went. Its position comes from the anchor-name the code already writes on the button, so the placement asks nothing of source.

Testing

• Everyday suite green on this branch: uv run pytest tests — 723 passed, 6 skipped.
• The five: test_suggestion_controls_stay_out_of_the_column, test_suggestions_sharing_a_block_keep_source_and_keyboard_order, test_a_settlement_in_a_reply_leaves_its_own_anchor_on_the_page, test_a_mark_in_the_layer_promises_no_press_the_layer_will_not_take, test_a_float_the_panel_displaces_hands_the_page_no_sideways_scroll — all pass with --run-nightly.
• Whole files: test_render_margin.py, test_render_conversations.py, test_render_widgets.py, test_render_outbox.py, test_render_navigation.py — the only failures left in them are ones that also fail on origin/main unchanged, at identical values.
• prettier --check and node --check pass on both JavaScript files; CSS is prettier-excluded here by .pre-commit-config.yaml.

How each was measured

The margin floor, on SUGGESTION_PAGE at 1200×900, before and after pressing Threads:

BEFORE  bodyRight 1185  main 768  --strip-r "max(0px, 250px, 0px, 0px)"  --claim-rail 250px
AFTER   bodyRight  765  main 765  --strip-r "max(0px, 0px, 0px, 0px)"    --claim-rail ""

The same probe at 335bf9a's parent reports --claim-rail: 250px in both postures, with data-lf-cramped standing — d226ee2 had already taken that veto off the rail claim, so the 900 floor was the window's question and nothing else.

The float, on LONG_PAGE at 1600×900 with the composer in the margin and the panel opened:

body.scrollWidth 1487   body.clientWidth 1165   composer.left 1166.9   composer.right 1486.9

The mark, on REPLY_TRAVEL_PAGE with an lf-options frozen into a reply: #tv-decision is present and checkVisibility() is true once the panel is open, its box 355×32, and .lf-mark-el is on nothing — at +6s as much as at +0. The same probe at 335bf9a's parent has the mark the moment the panel opens.

The tab walk from a suggestion's Accept, before: accept → reject → marker → preview close → preview action → preview action → next row's accept.

What this does not fix

Two of the four groups #127 recorded are still open, and both are a product decision rather than a stale reading.

• A covering panel is a modal dialog now, so the document behind it is inert: test_selection_hints_do_not_name_page_content_behind_a_covering_panel presses s at 700×900 and gets no hints because the page cannot hear the key. Whether a covering panel should take the page out of reach entirely, or cover it and leave it working, is the call — test_coarse_pointer_chrome_gives_its_compact_controls_humane_aims, test_workspaces_replace_each_other_instead_of_stacking[500] and test_coarse_pointer_resize_reach_stays_reachable_without_trapping_scroll fail as <dialog … class="lf-ui lf-panel open"> intercepts pointer events on the same fact.
• Escape belongs to whichever layer the browser is holding — the four test_render_anchors.py failures and test_esc_hands_the_page_back_after_it_has_closed_the_last_panel. Give the versions menu back its way in, its way out, and the reader's place #137 is open on this.

Separately, 0d7e0d0 "Open threads beside their source" landed while this was being verified and brought two of its own new tests red on main — test_the_shipped_long_thread_opens_beside_its_source_in_the_right_margin and test_the_full_thread_posture_follows_the_page_container_and_left_claims. They fail identically with and without this branch's changes; noted here so the next run does not re-derive them.

---

Automated fix for failed run

[leaf-agent]

Verified the claimed path on the merge ref: the five named tests pass with --run-nightly, and the five failures left in test_render_margin.py, test_render_widgets.py and test_render_navigation.py reproduce at 0d7e0d0 unchanged with the same assertions, so none of them is this branch's.

One finding, inline: skills/leaf/CLAUDE.md still carries the ownership sentence 335bf9a wrote alongside the code this PR restores, so the contract file and the runtime now disagree about who measures the shell.

[leaf-agent]

The gate on addfc26 went red on test_the_stop_hook_records_the_ending_of_the_turn_behind_a_claim — not this branch's, and not a browser test. It is a second-boundary race written into the test itself, so I reran the job (green on attempt 2) and am recording the mechanism here rather than opening an eleventh bot PR against a queue this deep.

The test's second Stop for s1 is allowed to move the stamp — >= closed — because close_turn writes now_iso() on every accepted call. The assertion after the s2 call then compares against closed, the value from before that permitted move. Cross a one-second boundary between the two s1 calls and the third assertion fails on a stamp s2 never touched, which is exactly what attempt 1 reported: assert '2026-08-30T20:36:19+00:00' == '2026-08-30T20:36:18+00:00'. The behaviour under test — another session's ending saying nothing about this claim — held; the reading of it did not. The fix is to re-read the stamp after the second s1 call and compare s2's result against that, so the assertion names the fact the comment claims.

Evidence

Attempt 1 of run 33333878596, job test, 1 failed, 722 passed, 6 skipped:

>       assert service_model.page_claim(claimed)["turn_closed"] == closed
E       AssertionError: assert '2026-08-30T20:36:19+00:00' == '2026-08-30T20:36:18+00:00'
tests/test_interact_session.py:2906: AssertionError

The branch does not touch tests/test_interact_session.py or anything under skills/leaf/scripts/, and test was green on the prior head 741fe2a (check run 99313101250). addfc26 changes skills/leaf/CLAUDE.md alone. Attempt 2 of the same run passed, and every gating check on addfc26 is green.