test: repair 7 failing test-all.sh files, fix ROLE_COLORS under a CB preset (#174 part B)

[dborup]

Summary

Part B of #174: repairs the seven failing test-all.sh files that are not covered by part A, one commit per file (1438 and 1446 together). Every failure had the same shape: a production change was intentional and the test was never updated, because none of these files runs in CI.

One of the failures was not a stale test. It exposed a real regression in public/roles.js (see "Real regression found").

Relates to #174

Per file

File	Cause	Choice	Coverage / what the test guards now
test-frontend-helpers.js	favStar() moved from the ★/☆ glyphs to a Phosphor sprite <svg> in Kpa-clawbot#1648 M2; the two assertions still grepped for the glyphs. The tbody.contains is not a function log (13 times) was a missing contains() on the nodes WS sandbox element; renderRows() has called it since Kpa-clawbot#1616. The log did not fail the run.	(a) update expectation + add stub	#ph-star-fill / #ph-star, the on class and aria-pressed
test-channel-psk-ux.js	channels.js uses ph-lock-open / ph-lock instead of 🔓/🔒 since Kpa-clawbot#1648 M2	(a) update expectation	asserts the badgeIcon mapping: user-added → ph-lock-open, encrypted → ph-lock
test-analytics-channels-integration.js	the "Channel Analytics" chip was dropped on purpose by the chat-app redesign (Kpa-clawbot#1376)	(a) invert the assertion	source guard that #/analytics / .ch-analytics-link stay out of channels.js; the rendered header is covered by test-issue-1367-channels-chat-app-e2e.js
test-issue-1648-m3-emoji-scan.js	the ping bot (2026-07-23) added 🏓 pong! (reply text) and a 🤖 avatar to channels.js after the M3 sweep	(a) narrow allowlist, 2 exact substrings	any other emoji in channels.js still fails. The 🤖 avatar is real iconography and is only exempted because the sprite has no robot symbol; TODO in the test to add ph-robot and drop the entry
test-issue-1438-customizer-mcrole.js	since Kpa-clawbot#1449, applyCSS removes the documentElement --mc-role-* write while a preset is active, so the preset colour only comes from body[data-cb-preset] in style.css. The test read the root inline style (now empty).	(a) model the cascade	effective value = body inline > body[data-cb-preset] rule parsed from style.css > root inline; server colour must still stay out of the root write
test-issue-1446-cb-preset-cascade.js	same cause (scenario 5 precondition)	(a) same helper	same
test-issue-1470-node-tile-helper.js	Kpa-clawbot#1533 renamed voyager-inverted to carto-voyager-dark and Kpa-clawbot#1614 made provider.url a lazy function. MC_setDarkTileProvider() returns false for an unknown id, so the default stayed active.	(a) new id, resolve lazy url, assert the setter accepted	selected provider, getTileUrl, _applyTilesToNodeMap url and invert filter

Nothing was deleted: no test was redundant.

Real regression found (separate commit, public/roles.js)

ROLE_COLORS is a live getter that reads --mc-role-* on documentElement. Since Kpa-clawbot#1449 an active colourblind preset is delivered only through body[data-cb-preset], which documentElement never sees. Under a preset, CSS-variable consumers showed the preset while every JS consumer of ROLE_COLORS (analytics, map, nodes, live, route views) kept the Wong default.

Observed in the browser with meshcore-cb-preset=deut, before the fix: body computes --mc-role-repeater #FE6100, documentElement #D55E00, ROLE_COLORS.repeater #D55E00. Applying prot at runtime did not change ROLE_COLORS either.

Fix: _readCssVar resolves on <body> (fallback documentElement). <body> inherits every :root value, so reads without a preset are unchanged. Regression test: new section G in test-issue-1407-cb-preset-propagation.js (red on the old roles.js). This commit is independent: drop it if you would rather handle it in its own PR; the test commits do not depend on it.

No production file on the do-not-touch list was changed. No real regression was found in channels.js.

Mutants (run in a copy of the tree)

Test	Mutant	Result
frontend-helpers	favStarIconHtml always ph-star	red (filled-star test)
frontend-helpers	favStarIconHtml always ph-star-fill	red (empty-star test)
frontend-helpers	remove the contains() stub	stays green: it only silences 13 log lines
channel-psk-ux	user-added badge switched to ph-lock	red
analytics-channels	re-add a #/analytics chip text to channels.js	red
1648-m3	extra emoji in the avatar / pong line	red
1438 / 1446	server colour written to --mc-role-* while a preset is active	1438 red, 1446 green
1438	style.css deut companion colour changed	1438 red
1446	style.css deut repeater colour changed	1446 red
1470	invert filter not applied / getTileUrl ignores provider / getActiveTileProvider returns null / voyager url wrong	red in each case
1407 §G	old roles.js	red

Results

test-all.sh file list (104 files), each file run on its own, 120 s limit, Node v26 locally.

	failing files
master 84c637ea	12 (exactly the list in #174)
this branch	5: test-observers-headings.js, test-issue-1648-m6-final-sweep.js, test-issue-1648-m6-lint-self.js, test-issue-1485-live-anim-z.js, test-naive-banner-tone.js (all part A)

No new failures. Every other test file that loads roles.js has the same exit code on master and on this branch, apart from the ones this PR repairs.

Also failing on master, not in test-all.sh, not touched here: test-map-clustering.js (makeClusterIcon pill assertion).

Browser (local server, e2e fixture DB, branch public/)

• Presets deut, prot, trit, achromat: body computes exactly the preset table values, and ROLE_COLORS now matches them; clearPreset() returns to the server palette.
• Dark mode: MC_setDarkTileProvider('carto-voyager-dark') returns true, getTileUrl() is the voyager url, getActiveTileProvider() carries the invert filter, the old id voyager-inverted is rejected. On a real node detail page the inset map loaded voyager tiles and the tile pane had the invert filter.
• Channel sidebar at 1280 px: a user-added PSK row renders ph-lock-open. At phone width the row renders ph-key (a separate mobile renderer, not changed here).

Not verified

• A server-encrypted channel row (ph-lock) in the browser: the fixture has no encrypted channels. Only the source mapping is tested.
• Presets were applied through MeshCorePresets in the page, not by clicking through the customizer panel.
• Node 22 (CI version); CI does not run these files today.
• No screenshots (browser pane).

🤖 Generated with Claude Code