Fit the Page map list inside its sheet and read six nightly fixtures at the layer's current contracts

[leaf-agent]

Problem

The nightly leg of ci failed with seven cases on run 33702158485 (c370784). All seven reproduced locally on every run, so none was the load-shaped race that running-tend warns about; each was a nightly fixture reading a contract #232 and #233 replaced, or a defect their new surfaces brought with them.

Three of the seven are no longer this branch's to answer. main reached them from the other side while this PR was open, and the merge below takes main's version of each:

• e71fa11 (Restore Page map chord grammar and focus) landed the Page map's Escape fix — the same leavePageMap route this branch had, under one flag (sheetCloseOwnsFocus) rather than two — and read its two fixtures. It also restored g m N as the numbered location prefix and g M as the complete sheet, reversing the grammar change this branch had adapted its fixtures to. So test_render_projection.py::test_escape_lets_go_of_the_ask_the_reader_is_standing_on reverts to main's g m 2, which is correct again.
• 991c64c fixed test_site.py::test_the_label_is_chrome_rather_than_words_to_quote by dropping the key-line reading rather than restating it.

Solution

What remains is one product fix and six fixture readings.

The sheet's list did not fit in the sheet. .lf-page-map-list was capped at calc(100vh - 150px) while the card it sits in stops at min(720px, 100vh - 24px). Measured on the corpus at 760×900: the sheet ran 172→892 and its list 270→1020 — 128px of the map hanging out of the rounded corner and off the bottom of the window, with its own focus ring drawn outside the card. The open sheet is a column now and the list takes what the head and search box leave: same measurement, 270→877.

Six fixtures read the layer as it was. #232's return stack means g L gives back the reading place rather than the tray's button, and a second g L into a standing tray is its own frame — two fixtures in test_render_controls.py. Three sibling widget-label drags in test_render_anchors.py read lf-pending straight off the mouseup, when the capture is queued behind it; all three now wait for the paint that says the gesture was read. And the ring walk's RING_NEW_STOP told "nothing focused" and "stood on this already" apart with one boolean, and the walk ended on either — a scope it joins part-way down its own order keeps its earlier stops on the far side of the tab order's crossing out of the document and back in, so ending there left page-map-search unwalked and unlit. Only a repeat ends a walk now; two empty reads in a row still stop one that never starts.

Testing

On the merge commit 91da9d3:

• uv run pytest tests — 808 passed, 6 skipped.
• The six fixtures this PR reads, plus both parameters of test_the_small_screen_map_is_a_complete_accessible_sheet and the two Page map fixtures the merge takes from main — all pass under --run-nightly.

The ring walk still fails, identically on main. test_every_ring_the_layer_draws_is_shown_whole_somewhere_in_the_corpus trips on a scope precondition — .lf-margin-marker[data-lf-kinds~="comment"] did not open a thread card on ship-review — before the loop this PR rewrites ever runs. Measured on both 91da9d3 and e71fa110 (main's tip at merge time), same assertion, same message. That cause is not this branch's; folding it in would put two unrelated causes through one review.

So merging this will not turn the nightly green on its own.

The seven original cases and where each went

case	cause	where it is answered
test_render_navigation.py::test_the_g_chord_reaches_panels_and_document_lists	the workspace restore needed the product fix	main, e71fa11
test_render_projection.py::test_escape_lets_go_of_the_ask_the_reader_is_standing_on	g m 2 — restored by e71fa11, so main's version is correct again	main, e71fa11
test_site.py::test_the_label_is_chrome_rather_than_words_to_quote	#232: c with nothing in hand names the page	main, 991c64c
test_render_controls.py::test_the_leaves_tray_takes_the_keyboard	#232: Escape returns the reading place, not the tray's button	here
test_render_controls.py::test_a_closed_leaf_clears_itself_off_the_tray	#232: two entries, two Escapes	here
test_render_controls.py::test_every_ring_the_layer_draws_is_shown_whole_somewhere_in_the_corpus	#233 added .lf-page-map-search:focus-visible; the walk ended at the tab order's crossing back through the document and never reached it	here, but blocked by the ship-review precondition above
test_render_anchors.py::test_a_widgets_label_takes_a_comment_inside_the_control_it_labels	the queued capture read a frame early — 26ms, deterministic under #233's heavier margin pass	here, with its two siblings

---

Automated fix for failed run

[leaf-agent]

This run was dispatched for the older run 33698883459 and reached the same seven cases independently. This PR covers all of them and is green, so I am not opening a competing one — its ring-walk change and the .lf-page-map-list overflow it found are both better than what I had. One divergence is worth a look before it lands, because merging pins it.

test_a_closed_leaf_clears_itself_off_the_tray now asserts two Escapes for the leaves tray, on the reading that the second g L is its own entry. But the reader is not in the tray when they press it. skills/leaf/CLAUDE.md says a frame "is active only while its owning surface still stands and the reader remains in the layer it entered", and navigation.panel.leaves only asks the first half — active: () => currentTray() === "leaves". So when the last row leaves on the poll, focus falls to body while the frame stays live, and the key line goes on offering esc back for a layer the reader is standing outside of. Measured on that fixture at the moment the row goes:

{'panel': True, 'expanded': 'true', 'rows': 0, 'focus': 'BODY', 'line': 'escback…'}

The second g L then stacks a frame whose captured workspace is the tray it found standing, so the first Escape restores the tray to itself and the second closes it. The test's own comment ("its own Escape gives that press back and leaves the workspace it found") is an accurate description of that, which is what makes it worth deciding rather than describing: it is one press giving back a workspace the reader never displaced.

I took the other branch — one entry, one frame — and it makes test_a_closed_leaf_clears_itself_off_the_tray pass unchanged, with the rest of this PR's readings intact:

The dispatcher-side alternative

--- a/skills/leaf/assets/runtime/keyboard/return-stack.js
+++ b/skills/leaf/assets/runtime/keyboard/return-stack.js
@@ -62,7 +62,15 @@ export function createReturnStack({ focused, paintHere, readingBlock }) {
       : null;
     const result = run();
     prune();
-    if (frame?.active()) frames.push({ ...frame, origin });
+    // One entry, one frame. A row whose surface the reader is already standing in has
+    // entered nothing: `g L` pressed inside the leaves tray only re-lands focus in it,
+    // and a second frame there makes one entry cost two Escapes — the first restoring
+    // the workspace that entry displaced, which is the tray still standing. The row
+    // names the entry, and the held frame must still be live: a surface the reader
+    // opened by pointer, or left and came back to, still gets its rung. `prune` trims
+    // only the tail, so a buried frame this row has already given up cannot block one.
+    if (frame?.active() && !frames.some((held) => held.id === row.id && held.active()))
+      frames.push({ ...frame, id: row.id, origin });
     return result;
   }

Full nightly on that branch, rebased on c370784: 1812 passed, 6 skipped in 1307.83s, with pre-commit run --all-files clean.

It is not the whole answer either — it leaves active() too loose, so the stale frame still advertises esc back from body until something pops it. Tightening navigation.panel.leaves to ask whether the reader is still inside the tray would make the second g L a real entry again and reach the same one-Escape result from the other side, which may be the reading you want.

One more thing from my own run, not this PR's: test_render_drafts.py::test_registered_control_keys_activate_once failed once in ~1800 tests with "the first press sent nothing" and passed in isolation, in its own file, and on a second full run. Its own comment says the DOM leads the log there; round_trip returns immediately when the attempt has not been registered yet, so the read can land before the post exists. Recording it here rather than filing it, since one observation is not an arrangement.

[leaf-agent]

Merged main (bf9dd47) into this branch to clear a conflict, pushed as 912c99d. Checks are green on that head.

The one conflict was test_site.py::test_the_label_is_chrome_rather_than_words_to_quote, and it needed a decision rather than a stitch. This branch updated the key-line assertion to the wording #232 gave it; 991c64c on main answered the same case by removing the reading altogether — it dropped says from AFTER_THE_DRAG and narrowed the test to the field-offered contract. That deletion landed on the non-conflicting hunk, so keeping this branch's assertions would have left them reading a key the dict no longer carries. I took main's version whole.

The Problem section still describes seven cases on the failing run; six of them are fixed here now, and the seventh is fixed on main. Updated the Solution and the case table to say so.

Resolution detail

Merge base c370784. Both sides rewrote the block after assert not label["fieldOffered"]:

• this branch: assert "comment on the page" in label["says"] plus a control["says"] control assertion;
• main: the block removed, and says: document.querySelector('.lf-keyline').textContent removed from AFTER_THE_DRAG — that hunk auto-merged, which is what made this branch's side unrunnable.

git diff origin/main HEAD after the merge is this PR's eight remaining files; git merge-tree against the live main is clean.

[leaf-agent]

Reviewed the merged tree at 912c99d. The change itself reads right and I could not fault it: the sheetReturning flag is set only on the dispatcher's own way out and consumed by the one close it queues, the sheet is the only surface that restores focus from a close handler (the keyboard reference does its restore synchronously inside setOpen, so the same task boundary does not bite there), and the test_site.py merge resolution is not just defensible but forced — bf9dd47 deleted says from AFTER_THE_DRAG, so carrying the branch's side would have raised KeyError on label["says"].

I ran the changed fixtures on the merged tree rather than reading them. The six that were red pass, plus both parameters of test_the_small_screen_map_is_a_complete_accessible_sheet. The ring walk I had to take apart, because it now fails before it reaches anything this PR changed — see below — but with ship-review held out it walks 540 stops with page-map-search lit and zero ring faults, which is both halves of the claim: the walk reaches the sheet's earlier stops through the crossing, and the sheet's list is inside the card that holds it (ring_faults is what would report the search box's ring drawn outside the sheet, and it reports nothing).

One thing the description no longer says truly, and it is worth knowing before merging rather than after. ## Testing is a nightly run on c370784; the base merge in 912c99d brought bf9dd47, and that commit reddened the nightly on its own. test_every_ring_the_layer_draws_is_shown_whole_somewhere_in_the_corpus now fails at .lf-margin-marker[data-lf-kinds~="comment"] did not open a thread card on ship-review, before the walk loop this PR rewrites ever runs — and it fails identically on bf9dd47 and on today's main tip e0a0549, so it is not this PR's. Merging this will not turn the nightly green: bf9dd47's own run leaves seven other cases red on top of the six here.

What I ran, and the split between the two causes

All on the merged tree 180441e (refs/pull/236/merge), host Chromium headless shell.

• tests/test_render_margin.py::test_the_small_screen_map_is_a_complete_accessible_sheet — 2 passed
• test_the_g_chord_reaches_panels_and_document_lists, test_escape_lets_go_of_the_ask_the_reader_is_standing_on, test_a_closed_leaf_clears_itself_off_the_tray, test_the_leaves_tray_takes_the_keyboard, test_a_widgets_label_takes_a_comment_inside_the_control_it_labels — 5 passed
• test_every_ring_the_layer_draws_is_shown_whole_somewhere_in_the_corpus — fails at ship-review; identical failure on bf9dd47 and on e0a0549. With ship-review removed from RING_WALK_EXAMPLES, assert not unseen, assert not errors and assert not faults all pass, stops is 540, and lit contains page-map-search. The remaining failure in that patched run is the expected one — the scopes whose only corpus was ship-review walked nothing.

Run 33714432132 on bf9dd47 is the whole picture. Six of its thirteen are the ones this PR reads — test_a_widgets_label_takes_a_comment_inside_the_control_it_labels, test_a_closed_leaf_clears_itself_off_the_tray, test_the_leaves_tray_takes_the_keyboard, test_the_g_chord_reaches_panels_and_document_lists, test_escape_lets_go_of_the_ask_the_reader_is_standing_on, and the ring walk's original cause. The seventh case in the description, test_site.py::test_the_label_is_chrome_rather_than_words_to_quote, is gone because bf9dd47 deleted the assertion rather than fixing it, which is why that hunk had nowhere to land. The other seven are bf9dd47's: test_every_passage_in_a_real_page_can_be_quoted[feature-gallery], test_ask_addresses_follow_the_feature_gallery_s_visible_margin_controls[1200] and [390], test_the_shipped_long_thread_opens_beside_its_source_in_the_right_margin, test_a_pick_states_the_whole_set, test_selected_reactions_keep_neutral_button_furniture[dark], test_a_comment_on_a_wrapped_diff_line_names_the_line_an_unwrapped_one_names, and the ring walk's new one.

Nothing here asks for a change to this PR — the ship-review card and the six beside it are a separate commit's to answer, and folding them in would put two unrelated causes through one review. The description's ## Testing line is the one thing worth restating against the current base.

[leaf-agent]

Restated ## Testing against the current base, which was the review's one ask. It now says which commit the green nightly was measured on (c370784, before the base merge), what is verified on 912c99d itself, and that merging this will not turn the nightly green.

The split between the two causes checks out from main's own run rather than from the merge ref alone: run 33714432132 on bf9dd47 — main without this branch — fails thirteen cases, six of them the ones this PR answers, and its ring-walk failure is already the ship-review thread-card precondition, not the walk-ends-early cause this PR fixes. So that seventh cause arrived with bf9dd47 and is unrelated to this branch. No code change here; the branch is unchanged at 912c99d.

[leaf-agent]

#239 takes the six nightly cases this branch does not, and one of them is the thread-card scope of test_every_ring_the_layer_draws_is_shown_whole_somewhere_in_the_corpus: ship review stands a contents map since bf9dd47, so its thread margin waits for 1472px of shell and the walk's 1440 opens Threads instead of a card.

Composed locally — this branch's head 912c99d with #239 applied on top — the walk gets past that scope, past page-map-search, and then stops on a fault neither branch's own run can reach:

in the page map sheet of corpus, the ring on div.lf-page-map-list "ask · Where should a ses" is not all there:
its top edge is 1.4px outside dialog.lf-ui.lf-page-map-sheet (ring 22,172,720,924 vs band 9,173,733,891)

The list runs 33px past the sheet's bottom and 1.4px over its top. Same box this branch refits, and the refit's measurement was taken at 760×900 while that stop is the 1200×900 walk. Left here rather than duplicated in #239, since .lf-page-map-list's sizing is this branch's.

[leaf-agent]

Two readings from run 33722764498 (30ed7c7) and a composition on it, both of which move this branch's ground.

e71fa11 on main reaches two of the six. It restores g m as the numbered prefix and g M for the complete sheet — the grammar this branch reads as gone — and preserves the workspace focus through the sheet's delayed close. Re-run on e71fa11 with nothing else applied, test_the_g_chord_reaches_panels_and_document_lists and test_escape_lets_go_of_the_ask_the_reader_is_standing_on both pass. The other five still fail as before. That commit also touches tests/test_render_navigation.py, and this branch no longer merges cleanly with it.

The page map sheet fault does not reproduce on a newer composition. main at 30ed7c7 with this branch and #239 (1e8d3868) applied walks 692 stops and reports one fault, and it is not .lf-page-map-list:

in the page of release-notes, the ring on div.lf-shotframe "Respond to shot · before"
is not all there: its top edge is under div.lf-shotrail "beforeafter"
(ring 21,627,664,898 vs 24,595,661,630, sampled 343,628)

Which of 7b70304, c9c2e80, e0a0549 or #239's second commit closed the earlier reading, I did not isolate — the claim here is only what the walk says on that composition. The lf-shot fault and the unnamed contents-link ring the walk finds after it are in #240, so the walk needs that branch as well as these two to finish.

[leaf-agent]

Reviewed the merged tree at 91da9d3. The base merge is what changed here: e71fa110 took over the Escape fix, so this branch is now the .lf-page-map-list refit, the six fixture readings, and the paragraph that explains the focus-return rule. The refit is the pattern the same stylesheet already uses for the keyboard reference — .lf-help.open { display: flex; flex-direction: column; } over .lf-help-results { min-height: 0; overflow-y: auto; } — and the [open] guard is needed for the reason the comment gives, since the author display would otherwise beat the UA's dialog:not([open]) rule. max-height on the sheet is already a border-box measurement (.lf-ui sets box-sizing: border-box), so capping the list against the box rather than the window is the right depth for it. The ring walk's third answer is sound: __lfSeen is reset immediately before the loop, so "seen" can never be the first read, and empty = 0 is only reachable after a stop was recorded.

One thing the merge left behind, inline. The paragraph was written against this branch's two-flag design — sheetActivation for the activation path, sheetReturning for the dispatcher's — and main collapsed both into the single sheetCloseOwnsFocus. Its closing sentence still describes the two-flag world.