fix: restore the separators the language file conversion dropped

[e107help]

Why

Fixes #6331.

31e7d1d3b06e012a2c299cb0f6b3d98d860d7dad ("Issue #5465 Core language files converted", 4 April 2025) rewrote the core language files from define() calls to returned arrays. 168 of the constants it moved carried a leading or trailing space, and not one of them came out with it. Where a call site leaned on that invisible space, the words now run together.

A visitor meets four of them. Who's Online reads Guests:3, Members:2; the search flood message reads Sorry, search is restricted to one search every30seconds.; a reply page titles itself Replying to:Hello everyone; and two subject lines leave the site altogether as Emailed item fromMy Site and Password reset fromMy Site. The rest are admin-side, mostly a label glued to the value it introduces.

The count in the issue needs correcting. #6331 says 62 call sites. It is 67, and the difference is in both directions:

• The audit dismissed two sites it should have kept out and five it should have caught. IMALAN_51 at image.php#L3811 and #L3933, LAN_UPDATE_23 at update_routines.php#L1218 and two neighbours, LAN_MAILOUT_241 at mailout_admin_class.php#L1402, and LANINS_038 at install.php#L1213 all lost a space a reader can see.
• The two required-field markers in the legacy user settings template are dropped instead. usersettings_template.php#L10 builds that marker as <span class='required'><!-- empty --></span>, and the only .required rule in the tree sits inside a commented-out block at e107.css#L116-L121 and was never more than a colour. Nothing visible sits beside the label, so nothing was lost.

What Changed

67 call sites across 23 files get a literal space in the concatenation. Eight commits, grouped by what a reader meets rather than by file.

The language files are deliberately untouched. A trailing space inside a translatable string is invisible to whoever translates it, gets trimmed by tooling, and every shipped pack would then need the same invisible space to stay correct. The separator belongs to the layout, so it goes in the layout.

A constant padded at the front gets its space in front of it: LAN_SEARCH_18, LAN_MAILOUT_70, LAN_MAILOUT_241, RL_LAN_057, EMOLAN_35, LAN_NOTWRITABLE and LANINS_038. EPL_ADLAN_72 was padded at both ends and gets one on each side.

Two things stay out on purpose:

• LAN_SEARCH_3, LAN_SEARCH_7 and LAN_SEARCH_8 and every site that reads them. Those are Search result detail lines and results caption are missing separator spaces #6298, and fix(search): separate result labels from the values they introduce #6325 is the open pull request for them. If that PR is closed or reworked, those sites come back into scope.
• Every site where the lost space never showed: inside a <td>, a <label> on its own, a title= or placeholder= attribute, an array value, a bcDefs map, immediately before a <br />, or where the string is used alone. Roughly a third of the references to these constants fall in that group, and The language-file conversion dropped the padding from 168 core strings #6331 lists them so nobody re-examines them.

How It Was Tested

Three tests, each proven to fail on the unfixed source by reverting it and re-running, never by reading:

• e107_tests/tests/unit/onlinePageTest.php renders online.php in a booted subprocess and asserts each count is separated from its label, including the most-ever-online line. It persists and restores the tracking preferences it needs rather than setting them after boot, where nothing would read them.
• e107pluginTest::testDependencyFailuresAreSpacedFromTheirValues calls XmlDependencies() with a missing plugin, a missing extension, an extension below its required version, an impossible PHP version and an impossible MySQL version. The extension-version case is there for EPL_ADLAN_72, the one constant in this change padded at both ends and so the only one that can be wrong in two directions.
• notifyTest::testUserVerificationLinesAreSpacedFromTheirValues captures the user verification notice and asserts all three lines are separated from their values.

The remaining sites have no practical test and were verified by reading, one at a time: the concatenation at the call site, against the constant's original padded value taken from the parent of 31e7d1d3b0. That covers the update log, the admin page messages and the search flood message. No claim is made that any of them was rendered in a browser.

The search flood message was given a page-rendering test and the test was then withdrawn. Driving that branch needs the flood interval, the visitor's own address and a row in tmp to line up, and the result passed alone but not inside the shuffled suite. An order-dependent test misleads everyone who meets it later, so the line is read-verified with the rest rather than covered by a test that cannot be trusted.

The whole unit suite (2732 tests), the parse check and the downgrade fixed point all pass on the branch.

Backwards Compatibility

Rendered HTML gains exactly one space per site, which is the defect being fixed. Nothing else about the markup changes: no tag, attribute or class list is touched, so a skin that has not been updated in a decade renders as it did before 31e7d1d3b0.

Shipped translation packs stay valid, because no language file changes. A pack translated after April 2025 carries unpadded strings and is now rendered correctly.

A pack that was never converted is worth stating plainly rather than leaving to be discovered. The loader still accepts an old-style define() pack and lets its values win, so on such a site the pack's own padding remains and the new literal space makes two. HTML collapses that everywhere except four plain-text sites: the three email subjects at fpw.php#L379, email.php#L182 and mail_manager_class.php#L971, and the notification body at notify_class.php#L342-L344. A recipient on an unconverted pack sees Password reset from My Site with a double space. That is the deliberate trade: a doubled space on a pack that has not caught up, against words run together on every pack that has. Putting the padding back in the language files would invert it, and that is the option being refused.

release/v2.3.x is not affected and must not be twinned. That branch never took 31e7d1d3b0, still ships "Guests: ", "Password reset from " and the rest, and renders correctly today. Applying this change there would double every space. Anyone tempted to finish the job should measure against that branch's own language pack first.

Sixteen of the changed lines are update log bodies, plus the ADMIN_01 and USET_15 entries. Core renders those bodies rather than parsing them, and no bundled plugin matches on them, so nothing in this tree reads the shape that changed. A third-party log shipper pattern-matching on a LAN_UPDATE_* prefix would see one extra space.

Third-party themes and plugins that concatenate these constants themselves still show the run-together text. Core cannot reach those, and the fix in each case is the same single space.

AI Model

Claude Opus 5 (claude-opus-5), as e107help.

Checklist

• The change is on master and does not belong on release/v2.3.x
• Rendered HTML changes only by the space each site lost
• No language file is modified
• Regression tests fail without the fix
• British spelling in prose and commit messages
• Every citation pinned to a commit