fix(alarms): strip only paired markdown, and three loose ends

[hokiepokedad2]

Four small things left behind by today's nine merges. One commit each so they read independently.

1. cleanRuleSummary stripped markdown characters that were not markdown — 73b554c

Flagged by #819. The old body was .replace(/[*_]/g, ''): every delimiter character, anywhere. Nothing in production descriptions contains one today, but **area names and saved-place labels routinely do** — work_gymwould have silently becomeworkgym` the first time Poracle rendered one into a sentence.

Stripping now runs in pairs: code spans first, then **bold**, *italic*, __underline__, _italic_. The underscore forms carry (?<!\w) / (?!\w) lookarounds, the same boundary Discord itself applies, so work_gym survives and a real _italic_ still loses its delimiters.

Checked against the live 5.2.1 server rather than assumed: all 19 descriptions across fort, invasion, lure, pokemon and quest use ** and nothing else, and every one survives the new function unchanged. processor/internal/rowtext/*.go agrees — ** is the only markup PoracleNG emits.

The underscore test was red before the fix (my_template → mytemplate); the paired-emphasis test is its legitimate-case twin.

2. The active-hours language-switch bug does not exist — 267e107

I reported this one and I was wrong, so this commit adds a guard rather than a fix.

The claim was that ActiveHoursChipComponent calls translate.instant() inside a computed(), which cannot see a language change. That was true on ngx-translate ≤17. This repo has been on v18 since #377, well before #816, and v18 backs its store with signals: instant() reads _currentLang and _translations, so calling it inside a computed() registers them as dependencies and a language switch invalidates the labels like any other signal read.

Verified twice — a probe counting recomputations across use('en') → use('it'), and the component's own test passing unmodified.

So the new test switches language through I18nService.use(), the path the language menu actually takes, and asserts the pill text goes from "hourly" to "ogni ora". It can be made red: wrapping the instant() call in untracked() — the shape a future ngx-translate dropping those signal reads would produce — fails it. A comment names the route so nobody files this again.

No changelog entry: nothing changed for users and there was no defect.

3. isStardust removed — 5864e63

Genuinely dead: usesRewardSlot took the template over in #821 and isStardust was read only by its own two assertions. Both assertions are kept, retargeted at usesRewardSlot — on a stardust rule that is the flag putting the floor field on screen, and in the pokémon-encounter case it is the flag stopping reward being overwritten on save. Same coverage, pointed at the property that is load-bearing. isPokecoins stays live; STARDUST stays because REWARD_SLOT_TYPES is built from it.

4. The [Unreleased] section had duplicate headings — e7df068

My fault, and worth fixing before a release rather than after. Squash-merging nine branches in a row left the section with two ### Added blocks and two ### Changed blocks: each merge resolved its CHANGELOG conflict by keeping both sides, which is right for the bullets and wrong for the heading above them. The release workflow promotes the section verbatim, so the published notes would have carried the duplicates.

All fifteen bullets are unchanged and still present; only the headings are merged, into Keep a Changelog's order.

Verification

npx jest --ci — 116 suites, 1392 tests (was 1389; three new). Production ng build succeeds. eslint and prettier --check clean on every file touched. No backend files changed. Seven files, +73/−14, no line-ending churn.

Noticed, out of scope

• prettier --check fails on CHANGELOG.md at baseline, before any change here. It sits outside the configured scope (src/**/*.{ts,html,scss}) so CI never sees it, but running prettier from the repo root trips on it.
• cleanRuleSummary's fort case is still poor — Fort updates: pokestop | distance: 5000m ["name"] keeps a raw JSON array mid-sentence, which is why fort cards do not render the line at all. Fixing it means translating change_types client-side, which is a feature rather than a cleanup.