feat(home): My Mesh "Node page" link (upstream 2027 port)

[dborup]

Summary

• Ports the My Mesh "Node page →" button from upstream Kpa-clawbot/CoreScope#2027 (reviewed upstream head e5536d20fa8ed4e94b7d2abec481fa382dbf11c7). It appears on the normal card (before Full health / View packets) and on the health-error card (as its only action), and navigates to #/nodes/<encodeURIComponent(pubkey)>. .mnc-actions now wraps on narrow cards.
• Local keyboard change, not part of upstream: the card's keydown handler called preventDefault() on Enter/Space before ignoring nested .mnc-btn / .mnc-remove targets, which suppressed the buttons' native keyboard activation. The handler now returns early for those targets, so Node page, Full health, View packets and Remove handle Enter/Space themselves without triggering the card's health action. Enter/Space on the card itself still opens health.
• Registers the self-contained regression test once in the existing "Run Playwright E2E tests" step, directly after test-home-coverage-e2e.js (no BASE_URL, CHROMIUM_REQUIRE=1).

Commits: 027045a1 (feature, keyboard change, test) and eb19b286 (test hardening, CI registration).
Files: public/home.js, public/home.css, test-issue-2027-my-mesh-node-page-e2e.js, .github/workflows/deploy.yml (+1 line).

Correction to the first commit message

027045a1's message overstates backend behaviour. The accurate statement is: the button navigates to the correct route; what the node detail page shows depends on the backend's node lookup (GET /api/nodes/{pubkey}).

• The mocked-404 test step is frontend evidence only (nodes.js renders its existing "Node not found" state for a 404). It does not show that all channel-only nodes are missing from the backend.
• UpsertNode is not the only writer to the nodes table; it is the only INSERT path, alongside several UPDATE/DELETE writers.
• In a local synthetic-backend diagnostic, a node moved to inactive_nodes still returned health 200 while node detail returned 404, so a normal card can also lead to "Node not found".

Regression test (24 scenarios)

test-issue-2027-my-mesh-node-page-e2e.js runs the real app.js, roles.js, home.js (and nodes.js for the 404 step) in headless Chromium with a mocked api():

• Card contents for normal and error cards; mouse, Enter and Space navigation for both; URL-encoding of the pubkey; Node page never triggers the card's health handler; Enter/Space on the card still opens health; Full health / View packets via mouse and keyboard; Remove via mouse, Enter and Space (keyboard cases assert exactly one removal, no navigation, no health panel); wrapping at 320px without horizontal overflow; re-render without doubled handlers; mocked 404; combined interactions.
• Any pageerror or console.error on a step's pages fails that step. Playwright's default timeout is 5 s. Exit 0 requires all 24 scenarios to run with zero failures.
• Playwright's own SIGINT/SIGTERM/SIGHUP handlers are disabled for this test, because they only close the browser and could let an interrupted run exit 0 without a summary. The test instead prints a FAIL — interrupted line (counted as a failure by scripts/aggregate-e2e-pass.sh), closes the browser and exits 128+signal.

Local verification (not GitHub CI)

Fresh, on an isolated export of the merge result of this branch onto current master 24760c3db917e908796045eeb0c4c5a3f98d0bc9 (after #53); merge tree 4721d3a1b7819bac002420b1988251e730b9d03b; macOS, Playwright 1.58.2 Chromium:

• Merge is conflict-free. Only the four files above differ from master. fix(frontend): handle orphaned api cleanup promise rejection #53's public/app.js, test-app-api-inflight-cleanup-rejection.js and test-all.sh are unchanged, and the workflow YAML equals master's apart from the one added line.
• CHROMIUM_REQUIRE=1 node test-issue-2027-my-mesh-node-page-e2e.js: 2 runs, 24/24, exit 0, no leftover browser or test processes.
• node --unhandled-rejections=strict test-app-api-inflight-cleanup-rejection.js: 8/8, exit 0.
• node --check, YAML parse, git diff --check and the three TestForkGuard* tests pass.

Earlier local runs on this branch (before #53 was merged; not repeated for this PR):

• 3 normal runs 24/24. With only the keydown guard removed, the 8 keyboard scenarios fail (Node page ×4, Full health ×2, Remove ×2) and 16 pass.
• SIGTERM / SIGINT / SIGHUP during an unfinished run: exit 143 / 130 / 129 with the FAIL marker, no summary, no leftover processes. The previous version of the test exited 0 in the same situation.
• Injected console.error and late pageerror were detected; a doubled Remove re-render was detected; a fatal exception after launch exited 1; CHROMIUM_REQUIRE=1 without Chromium exited 1; the registration line under bash -eo pipefail with tee kept failures non-zero.
• Independent review approved the last commit. Signal probes during chromium.launch() (0–600 ms) and during the final browser.close() (0–120 ms) found no exit 0 combined with a FAIL marker.

Known limitations

• Staging has not been tested.
• No GitHub CI result is claimed here; see the checks on this PR.
• test-home-coverage-e2e.js (existing test, needs a running server) failed its "Full health" step once in 4 local runs on the base without this change; with this change it passed 4 of 4. Treated as a pre-existing flake; not investigated.
• On the normal path, page.close() and the final browser.close() have no local timeout; they are bounded only by the job timeout (deploy.yml sets none, so GitHub's default applies). After a signal the test forces exit within 5 s.
• If a signal arrives exactly during the final browser.close(), the exit code relies on Playwright's close ordering rather than an explicit check (probes found no false pass). A rejection from page.close() during an interrupt can exit 1 instead of 128+signal; both are non-zero.
• The test is not added to test-all.sh.

🤖 Generated with Claude Code