test: test-all.sh fails 12/104 files on master and stops at the first one; none of them run in CI

[dborup]

Summary

test-all.sh is the documented way to run the unit tests (AGENTS.md, npm test), but on master 12 of its 104 files fail. The script uses set -e, so it stops at the first failure, which is test-frontend-helpers.js on line 18, and the remaining ~86 files never run in a normal invocation. None of the 12 failing files runs in CI (deploy.yml does not call test-all.sh and does not list these files), so nothing catches them.

Every agent works around this by running files one at a time and comparing against master. That is slow, and it hides real regressions behind "known baseline failures".

Found during #170 (PR #173). The list below was reproduced independently on master.

Failing files on master 84c637ea

Each file was run on its own from a git archive export of master, with Node v26 locally (CI uses Node 22). Each run had a 120 s limit.

File	First failure
test-frontend-helpers.js	the two known favStar failures (705 passed / 2 failed)
test-channel-psk-ux.js	user-added rows use a distinct unlocked icon (🔓) from server-encrypted (🔒)
test-analytics-channels-integration.js	channels.js sidebar links to #/analytics with "Channel Analytics" text
test-observers-headings.js	expected headings present and ordered
test-issue-1648-m3-emoji-scan.js	channels.js has 2 emoji/misc-icon hits
test-issue-1648-m6-final-sweep.js	5 emoji-as-icon violations
test-issue-1648-m6-lint-self.js	"repo MUST be clean before anti-tautology probe". Possibly an artefact of running outside a git checkout; verify in a real worktree.
test-issue-1438-customizer-mcrole.js	--mc-role-companion does not reflect the active preset (got empty)
test-issue-1446-cb-preset-cascade.js	precondition: deut active → --mc-role-repeater = #FE6100 (got empty)
test-issue-1470-node-tile-helper.js	getActiveTileProvider in dark mode: provider url voyager, got undefined
test-issue-1485-live-anim-z.js	animLayer hosts ≥3 .addTo() animation shapes (got 2)
test-naive-banner-tone.js	TypeError: document.addEventListener is not a function (sandbox stub missing)

Some of these were already reported as baseline in other PRs: favStar, channel-psk-ux, analytics-channels-integration, 1648-m3. test-live-anims.js, which is not in test-all.sh, fails too, for the same reason as 1485.

Likely causes (to verify per file)

• Stale source-text assertions. Fork changes to the UI (for example the Phosphor icon migration, channel sidebar changes and live animation refactors) were made without updating tests that grep the source code.
• Incomplete vm or DOM stubs for code that now uses newer browser APIs (tbody.contains, document.addEventListener).
• Customizer/preset tests that depend on CSS variables the test sandbox never sets.

Proposed fix

For each file, pick one of these and record the reason in the PR:

1. Fix the test. Update the expectation or stub when the behaviour change was intentional. Show that the test still catches a mutant of the behaviour it guards.
2. Fix the code. Do this when the test found a real regression, and add it to the PR description as a bug.
3. Delete the test. Only when it is redundant, and name the test that covers the same thing.

Then:

• Make test-all.sh pass end-to-end on master.
• Add a CI step that runs test-all.sh, or register the missing files in the JS unit step, so this cannot drift again. Fork guards stay unchanged: 9 × github.repository == 'Kpa-clawbot/CoreScope'.
• Consider printing a summary from test-all.sh instead of aborting on the first failure, while still exiting non-zero.

Splitting the work into a few PRs is fine, for example one for the 1648 icon tests, one for customizer/preset and one for the rest.

Acceptance criteria

• sh test-all.sh exits 0 on master after the fix(es).
• test-all.sh (or the equivalent file list) runs in CI.
• Every kept test that was changed has a mutant that turns it red.
• No new npm dependencies.

Relates to #170, #173.

[dborup]

Part A merged in #177 (d2dbe84e) and part B in #178 (470033ea). On master the 12 files that failed in test-all.sh now pass when each file is run separately (the reviewer measured: master 12 failing → both merged 0). #178 also fixed a real regression: since #1449, ROLE_COLORS ignored an active colour-blind preset (public/roles.js).

This issue stays open until test-all.sh runs in CI with a summary instead of stopping at the first failure. That follow-up PR is in progress. It also covers the review nits from #173, #177 and #178, and an inventory of about 70 root test-*.js files that neither test-all.sh nor deploy.yml runs.

[dborup]

Fixed by #177 (d2dbe84e), #178 (470033ea) and #187 (c94a0063).

• The 12 failing test-all.sh files are repaired. test: repair 7 failing test-all.sh files, fix ROLE_COLORS under a CB preset (#174 part B) #178 also fixed a real ROLE_COLORS colour-blind preset regression.
• test-all.sh now runs every file and prints a summary. It exits 1 if any file fails.
• CI's JS unit step runs only sh test-all.sh: 200 files, about 12 s on Node 22.
• 45 orphan root tests are registered and 3 more repaired.
• A guard (test-test-all.js) fails if a root test-*.js runs nowhere.

Leftovers are tracked in #189: 13 red orphan unit tests and 4 orphan E2E files, the collapsed-group caret, and the #187 review nits.