Read the page-map chord's progress from its key faces, not the key line's text

[leaf-agent]

Problem

ci on aa41f7a ("Keep chord hints stable through progress") failed with 33 nightly tests red. Most have since been fixed on main (#134, #137, #145, #148, #149, #150, #151, #152, #153) or are carried by an open PR (#155's ring walk, #156's catalog digest). The two still failing on main at c0a5b65 are test_render_margin.py::test_g_addresses_the_page_map_prefix_in_its_announced_order and test_one_margin_item_owns_a_targets_controls_information_and_more_actions, which #154's body already named as aa41f7a's and left for this run.

Both are stale readings of the key line, not a regression. They pressed g, asserted the line contained m page-map items, pressed m, then asserted it contained 1–9 page-map items — a text read that only distinguished the two stages while the line dropped the keys the chord had consumed. aa41f7a is the change that stopped it doing that: keyline.js now draws every destination's complete route (completeChordSteps) and marks progress with data-lf-key-state, so the row reads gm1–9page-map items both before and after m — which is the behavior the commit was for. That commit updated the same assertions in test_render_navigation.py and test_render_anchors.py; test_render_margin.py spelled them differently and was missed.

Solution

Assert what now carries the distinction: the row's keycaps and their faces. After g the page-map row reads g m 1–9 with the list key neutral; after m the same three keys, with that key pressed. This is the idiom aa41f7a established in test_render_navigation.py, and it proves more than the old read did — the route's identity and its stability under progress, not just that some substring moved.

The selector for one binding's keycaps is now key_line_route in tests/render_harness.py beside key_line, so the next change to this shape has one name to grep rather than a per-file spelling. The five existing sites in test_render_navigation.py are left as they are: three of them also need the enclosing .lf-key-sequence element for sequence_geometry, so converting them is a refactor of green tests rather than part of this fix.

Testing

On this branch at c0a5b65, with Playwright driving the real headless shell:

• uv run pytest tests/test_render_margin.py --run-nightly -n0 — 22 passed. Before the change, on the same checkout, both tests fail with Actual value: gTThreads panelgAAsks panelgMPage mapgm1–4page-map itemsgg / Gtop / bottomesccancel?more, which is where the 1–N between m and the word comes from.
• uv run pytest tests/test_render_navigation.py tests/test_render_margin.py --run-nightly -n2 — 84 passed. The navigation file is included because it owns the assertions this one is being brought into line with, and because the new helper lives in the shared harness.
• uv run pytest tests — 740 passed, 6 skipped. That is the gate CI runs on this PR.
• ruff check tests/ and ruff format over the changed files — clean.

Seventeen of the run's other failures were re-run here to place them rather than assume the merges cleared them: tests/test_render_drafts.py --run-nightly -n2 — 50 passed, so all fourteen Locator.fill timeouts are gone, and the three navigation failures (test_a_questions_digits_are_drawn_whole, test_c_in_a_seated_conversation_reaches_the_thread_it_is_in, test_the_ring_holds_on_a_seat_the_agent_has_still_to_answer) passed in the navigation run above. The rest sit in test_render_controls.py, test_render_options.py, test_render_projection.py, test_render_gate.py, test_render_conversations.py, and test_render_anchors.py; #155 already owns the one known to survive there.

---

Automated fix for failed run

[leaf-agent]

This and #157 are the same fix, opened twelve minutes apart by two ci-fix runs, and neither references the other — so whichever lands second conflicts rather than becoming a no-op. Only one should merge. I was dispatched for run 33344080789 (21332fc), derived this same change a third time, and am dropping it rather than opening a fourth.

Both edit the same two page.keyboard.press("g") sites in tests/test_render_margin.py onto the same reading — locate the row by data-lf-commands~="navigation.page-map-item", assert the keycaps spell g m 1–n, and take the progress off data-lf-key-state. The differences are where the helper lives and how the span is spelled: this PR puts key_line_route in render_harness.py beside key_line, and writes 1–9 out at the first call site; #157 keeps expect_page_map_address local to the test module and derives the span from the list's count through address.js's own 1–n-or-1 rule, so a corpus change that drops PAGE_MAP_PAGE to a single map location keeps that site honest. Against that, the harness is the better home for the selector if these five sites in test_render_navigation.py are ever converted, which this PR's body already notes it is deliberately not doing.

Both are correct as tests: the row's text is unchanged by the m press, so the settle has to rest on a positive fact, and in both it does — to_have_text only matches once g has armed the chord and drawn the row, and the post-m assertion is the retrying pressed read rather than the neutral one.

Verification

uv run pytest --run-nightly over all 32 node ids from run 33344080789 at c0a5b65: 3 failed, 29 passed — these two margin cases and #156's catalog digest. With the equivalent change applied locally, tests/test_render_margin.py --run-nightly — 22 passed, and the full nightly surface green on this runner: everyday tests 740 passed / 6 skipped, the other nineteen test_render_*.py files 780 passed across three batches, test_site.py 20 passed.

[leaf-agent]

Closing this as the duplicate, in favour of #157 — the comment above established that only one should merge, and nothing since has picked one, so the second to land would conflict rather than no-op.

#157 is the one to keep on three counts: it was opened twelve minutes earlier, its test and lint checks are green at a9411f8 while this branch's are still running, and it already carries a review. On the substance, it derives the span from the list's own count at both call sites through address.js's 1–n-or-1 rule, where this PR writes 1–9 out at the first site — the difference the comment above already named as the honest one.

The one thing worth keeping from here is key_line_route in tests/render_harness.py: a single named selector for "the keycaps one key-line binding draws" is the better home if the five sites in test_render_navigation.py are ever converted, and it is a small graft onto #157 if a maintainer wants it. The branch is left in place rather than deleted, so nothing here is lost.

Diagnosed while handling run 33346442884 (6d68a1b), whose five failures resolve to #157's two margin cases, #156's catalog digest, and two Asks (N) count cases that #151 already fixed on main.