Fix pytest suite (main CI) + BEM landing page restructure

[diese-tech]

Summary

Two independent changes ended up on the same branch/PR — noting that upfront since they're unrelated:

1. Fix main's CI (urgent)

main has been red since PR #14 merged. Root cause: PR #13 (the pytest suite) and PR #14 (block list / live panel state / admin-override logging) were developed in parallel against the same pre-merge base, coordinated to avoid touching the same files — which they didn't — but #13's tests still encoded the API surface and behavior exactly as it stood before #14 changed it. The moment both merged, 9 of the 92 tests failed against the new, intentional behavior.

This fixes the tests only — no production code changes. The behavior encoded by PR #14 (already reviewed, including two Codex-flagged fixes) is correct; the tests were stale.

What was stale and why:

• test_command_tree.py — EXPECTED_COMMAND_NAMES didn't include yap block, yap unblock, yap modlog (all added in Add block list, live panel state, and admin-override audit logging #14).
• test_panel.py — refresh_panel_message dropped its owner parameter in Add block list, live panel state, and admin-override audit logging #14 (it now derives owner, lock/hide state, and the permit list from the DB/channel itself). Updated all four call sites from 3 args to 2, added a test for the owner-left-guild no-op path.
• test_room_actions.py — apply_transfer/apply_kick now call bot.storage.get_active_temp_channel (needed for the new log_admin_action), missing from the shared _bot() fixture. Updated two refresh_panel_message call-arity assertions from 3 args to 2. Added a test locking in that an admin-initiated transfer is logged against the pre-transfer owner, not the post-transfer one.
• test_temp_channels.py — the reconcile-time panel backfill now posts the room's real locked/hidden/permitted state instead of defaults; added a dedicated test covering an already-locked/hidden/permitted room.

pytest: 95 passed, 0 failed locally (was 9 failed, 83 passed, reproducing the actual CI failure first).

2. Restructure the landing page to BEM naming

Renamed every CSS class in docs/index.html to block__element--modifier form and reorganized the previously single-line minified stylesheet into commented sections, one per block, so the page is easy to hand-edit going forward.

Blocks: page-shell, site-header, button (+button-row), hero, stats, showcase, card (generic, reused by the features grid / steps / two-column sections / closing CTA via --accent-a/--accent-b/--cta modifiers), page-section (+--split), features, steps, code-sample, permissions, site-footer. .muted became the utility class .u-muted. The two positional nth-child shadow variants on the feature cards became explicit card--accent-a/card--accent-b modifiers, so reordering the cards in HTML no longer silently changes which one gets which shadow.

Purely a rename/reformat — verified pixel-identical rendering via a headless-Chromium screenshot before and after (byte-identical PNG, same SHA256) at 1400×1000. No visual, structural, or script change; the stats.json fetch script and its target element IDs are untouched.

Test plan

• pytest: 95 passed, 0 failed
• py_compile clean
• No production .py file touched
• Landing page: before/after screenshots byte-identical (same SHA256) — zero visual regression