fix(ui): refresh the open help panel when the language changes

[argszero]

Summary

The shortcut help panel (#help-panel) is a non-modal floating panel: it does not cover the
Settings language select, so the user can change the language while it is open. Its heading and
close button carry [data-i18n] and are rewritten by applyStatic(), but the four shortcut rows
and the #help-context line are built by renderHelp() — whose only call site is toggleHelp's
open branch. After a language switch the same panel therefore shows two languages at once, and
nothing short of closing and re-opening it redraws it.

atp:langchange already re-renders the identically situated tour popup (renderTourStep()); this
adds the help panel next to it. The guard reads classList.contains("hidden") — the mechanism
toggleHelp hides the panel with. The property form (!$("#help-panel").hidden) is always false
here, so it would turn the branch into an unconditional repaint.

Related Issue

No linked issue: the repository has no open issue for this. Reported here with reproduction.

Changes

• ui/js/app.js: the atp:langchange handler re-renders #help-panel when it is open
• src/state_gate.rs: new rule the_language_switch_refreshes_every_overlay_it_can_show,
  which derives what must follow the language switch instead of restating it:
  the overlay set from ui/index.html (top-level elements after #app carrying the hidden
  class), modal-vs-non-modal from ui/css/style.css (a class rule carrying inset: 0 — a modal
  overlay blocks the language select, so it cannot be open when the user switches), and the
  writers from ui/js/app.js (a function that writes one of the overlay's ids and calls T().
  A future overlay or a future writer nobody refreshes turns it red. The four companion tests
  (_roster_is_real, _rules_have_teeth, _rules_separate_the_variants,
  _fixed_guard_is_the_edit_sheet_text) give each rule its own teeth and declare what happens
  to the competing fixes.
• ui/README.md: the refresh list gains the help panel, plus a bullet recording the invariant
  and the gate's scope
• ui/index.html: cache-bust js/app.js?v=20260922-10 → -11
• no config/data-structure change

Tests

• cargo test — 361 passed / 0 failed (baseline before this branch: 356)
• cargo fmt --check — clean (rc=0)
• new unit tests added (5: the rule, its roster, its teeth, the variant relations, the edit-sheet
  sync)

Checklist

• branch name follows the convention (fix/…)
• commit message uses Conventional Commits
• single responsibility, minimal diff

Evidence

Two instruments, one for the shape and one for the fact.

jsdom probe (r94_probe.js, real app.js in real index.html, real ? / language-change
events; the language-discriminating values are read out of window.I18N, not hardcoded):

tree	variant	axis-fail legs
pre-fix (6036d759…)	base	A1, A2, Z1 — the panel keeps the old language
pre-fix	fix	none
pre-fix	m_close (competing fix: close the panel)	A4 — the switch steals the panel the user opened
pre-fix	m_datai18n (competing fix: hang data-i18n on the JS rows)	A2, Z1 — a template with {view} cannot be maintained that way
shipped bytes (c6ca9310…)	landed	none — 8/8 legs

shape gate (r94_langchange_roster_scan.py, rosters derived from ui/index.html +
ui/js/app.js + ui/css/style.css): pre-fix RED: R1:help-panel; with the fix ALL GREEN;
m_close RED: R1:help-panel (closing the panel is not refreshing it); shipped bytes ALL GREEN.

compiled gate (r94_compile_gate.py, splices the rule into the real src/state_gate.rs,
rustc + clippy + tests): 3 legs (base / fix / live), 30/30 as declared — base red on
exactly the_language_switch_refreshes_every_overlay_it_can_show, fix and live green; the
live leg's app.js md5 is the shipped c6ca9310…, so the tree under test is the tree merged.
Negative control (rule 1 disarmed): every leg red on exactly {_roster_is_real, _rules_have_teeth, _rules_separate_the_variants}, and the unfixed tree's axis test turns green — it was red
because of rule 1, which is the direction that is easy to get backwards.

Scope, stated honestly: the rule is lexical. It proves the refresh roster covers every writer of
every derived non-modal overlay; it does not prove that what is on screen is in the current
language at that instant (that half belongs to the jsdom probe), and it cannot see the competing
"just close the panel" fix — the probe's A4 leg rejects that one. Recorded in ui/README.md.

[argszero]

Self-review (the PR author is this account; --approve is rejected on one's own PR, so this is a plain comment, not a formal review).

What I checked before merging

• The guard's mechanism is the one the element itself uses: toggleHelp(false) hides #help-panel with classList.add("hidden"), so the branch reads classList.contains("hidden"). The sibling-property form would be a constant-true guard and an unconditional repaint — the probe's D2/E2 legs reject it.
• The insertion point is the handler that already refreshes the identically situated tour popup, not a new listener: the file still has exactly one atp:langchange listener (counted before and after).
• renderHelp gained exactly one call site (derived from the live count, not asserted). No new i18n key: T("help.* count is unchanged, and the fix uses no literal text.
• The gate is derivation-based, not a snapshot: the overlay roster, the modal/non-modal split and the writer roster are all computed from ui/index.html / ui/css/style.css / ui/js/app.js. There is no exemption list — #chat-modal is excluded because its class rule carries inset: 0, a fact read out of the stylesheet.
• The declared-vs-actual tables are printed side by side, and the m_close boundary is measured rather than hidden: both instruments reject "close the panel on a language switch", but for different reasons — the gate because renderHelp was not called (R1 asks whether the writer ran), the probe because the user's panel was taken away (A4). That relation is asserted in the_r94_rules_separate_the_variants.
• The merged tree is the tested tree: the live leg's app.js md5 (c6ca931073951cc400245036850f3133) is the one in this branch, and the gate splices the rule into the real src/state_gate.rs, so rustc + clippy -D warnings + the test run all measure the shipped text.

Honest limits

The rule is lexical. It shows the refresh roster covers every writer of every derived non-modal overlay; it does not show that the pixels are in the current language at that moment — the jsdom probe owns that half — and it cannot see a fix that removes the panel instead of repainting it. Both limits are recorded in ui/README.md.

While landing I found two bugs in my own instruments, both of the "trip wire was wired to the wrong fact" kind, and fixed them here rather than papering over the readings: the applier's read-back leg asked for a marker comment the splice intentionally drops, and the gate's E1-direction leg demanded the pre-landing XOR, which reports a false red on an already-landed tree (an "insert after an anchor" edit keeps old x1 in both states, so the state has to be read off new).