[rumblefrog]

Summary

Fixes #1448 — the mobile-viewport burger button rendered with the user-agent default grey 1px-border chrome instead of the panel's ghost-button styling.

The button in web/themes/default/core/title.tpl shipped class="btn--ghost btn--icon" — modifier-only, no base btn. In theme.css:

• .btn is the load-bearing rule: it declares the --btn-bg / --btn-color / --btn-border defaults AND applies them as background / color / border / display: inline-flex / padding / height.
• .btn--ghost / .btn--primary / .btn--secondary / .btn--danger are pure custom-property overrides — they SET the variables but don't apply them.
• .btn--sm / .btn--icon / .btn--xs layer geometry on top but still don't carry the chrome declarations.

Without .btn in the chain, the variables get set but nothing reads them — <button> falls back to UA-default chrome (visible grey pill that fights the topbar's dark background on mobile, where [data-mobile-menu] flips from display: none to display: inline-flex at <=1024px).

Sister sites swept in the same PR

• core/footer.tpl — the palette Esc close button shipped the same modifier-only chain. Same live wrong-chrome shape behind the <kbd>Esc</kbd> child.
• partials/player-drawer.tpl — drawer-close X button. Reference template, not live render path — the live drawer chrome is rendered by web/themes/default/js/theme.js's renderDrawerBody() / renderDrawerLoading() (which always shipped the correct three-class chain). Documentation sync, not a behaviour fix.

Regression guard

New test web/tests/integration/ButtonClassChainTest.php — single-method parser-style sweep across:

• web/themes/ (every Smarty template + the install wizard chrome)
• web/themes/default/js/ (the runtime HTML emitter)
• web/scripts/ (panel-side JS)
• web/pages/ (page handlers)
• web/includes/View/ (view DTOs)
• web/install/, web/updater/, web/api/handlers/ (preventive coverage for any future inline-HTML emission)

For *.tpl, *.php, AND *.js. Every class="…" / class='…' attribute body carrying a btn--* modifier MUST also carry the base btn token. The JS-side coverage is what makes the gate's "panel chrome" protection structural — theme.js ships seven runtime <button class="btn btn--ghost btn--icon[ btn--xs]"> strings via template concatenation; without scanning them the gate would paper over the live bug surface.

Implementation details:

• Strips Smarty {* *} comments (default + -{* *}- non-default delimiters used by page_login.tpl / page_blockit.tpl / page_kickit.tpl / page_admin_servers_rcon.tpl), PHP comments via php_strip_whitespace(), and JS /* */ + // comments so AGENTS.md-style explanatory blocks don't false-match.
• Negative lookbehind (?<![-_\w]) on the class keyword skips data-class= / aria-class= siblings.
• Sanity-check assertions (>100 files scanned, >200 attribute bodies extracted) guard against silent zero-scan if ROOT / scanRoots() ever drift.
• Smarty-conditional class chains (anything containing {) are skipped; false negatives bounded to chains with NO base in any branch (none in the codebase today).

I deliberately introduced a regression in theme.js during development to verify the gate fires; it caught the bug at the right line with the right message.

AGENTS.md changes

• New "Anti-patterns" entry documenting the bug class with the user-reported regression details, sister-site rationale, and regression-guard scope.
• New "Button class chains" Conventions block (positive-shape contract paired with the anti-pattern).
• New "Where to find what" row pointing at the regression guard.

Adversarial review trail

Spawned an adversarial reviewer subagent on the initial diff. It surfaced:

• theme.js runtime-emitted button strings were outside the scan scope (MAJOR — fixed by extending scanRoots + adding JS comment stripping).
• The "PURE CSS-variable overrides" claim was technically wrong for sizing modifiers (MAJOR — tightened wording across all five sites).
• Missing "Where to find what" row (MAJOR — added).
• The drawer-template framing as a "live sister bug" was misleading (MAJOR — clarified in the AGENTS.md entry; the JS scan extension also makes the framing structurally honest now since the gate covers both halves).
• Several MINOR / NIT findings around docstring honesty, scan scope completeness, regex edge cases, and template-comment bloat — all addressed in this commit.

Test plan

• ./sbpp.sh phpstan passes (level 5, baseline unchanged)
• ./sbpp.sh ts-check passes
• ./sbpp.sh composer api-contract is in sync (no diff)
• ./sbpp.sh test — full suite passes (891 tests, 3338 assertions, including the 3-assertion regression guard)
• Test catches a fresh regression — verified by intentionally breaking theme.js's renderDrawerLoading() button, watching the gate fire, then reverting
• Visual verification via Playwright on mobile viewport (iPhone 13-like 390×844, light + dark themes, before-and-after screenshots) — burger button now renders with the panel ghost-button chrome instead of UA-default grey

CLA: I have read the CLA Document and I hereby sign the CLA