fix(de): correct Gemma 4b mistranslations in Localizable.strings

[michi-onl]

Fix 9 incorrect German translations introduced by automated Gemma 4b translation in Stats/Supporting Files/de.lproj/Localizable.strings.

Changes

Key	Before	After	Reason
Number of s-cores	%0 Super-Prozessoren	%0 S-Kerne	Pattern consistency with e-cores/p-cores
Box	Karton	Box	Established UI loanword
Tachometer widget	Drehzahlmesser	Tachometer	Avoids automotive jargon
Load per core	Laden pro Kern	Last pro Kern	Matches existing "Last" usage in file
Efficiency core	Kern der Effizienz	Effizienzkern	Matches existing "Effizienzkerne"
Performance core	Leistungszentrum	Leistungskern	Matches existing "Leistungskerne"
Super core	Hervorragende Kernfunktion	Superkern	Semantic error
Upload speed	Hochladen	Uploadgeschwindigkeit	"speed" was missing entirely
Download speed	Herunterladen	Downloadgeschwindigkeit	"speed" was missing entirely

[michi-onl]

Note to the maintainer: Using a SKILL.md file together with a 30B+ model would lead to significantly better results.

[exelban]

I would like to run bigger model but have not enough vram for that)

[michi-onl]

How were these Gemma translations accomplished? I believe a combination of a good cli harness, full file context, and a skill.md file would be necessary. If these resources are not available, we would need to wait until the models improve or the workflow becomes more accessible.

[exelban]

basically I just wanted to fix the problem that a lot of languages has no translations for new keys at all. So even not best translation will be better than no translation of some words at all. I don't want to spend a lot of times for that, but if you want you can help improve that. The translation workflow could be found here.

[michi-onl]

Existing set-and-forget solutions aren't compatible with swift project currently. Despite optimizing the script's context and prompts, local models under 8B parameters failed to follow instructions or output valid structured data with good context awareness. Reliable translations imho currently require a frontier AI model. If cost is a concern, platforms like OpenCode offer limited free access to capable models.

I'll look into improving other locales.