Deploy Gemma 4 E4B + G1 prompt; remove Swahili (English-only)

[nmrenyi]

Summary

Deploys Gemma 4 E4B + the G1 deflection-fix prompt, and removes Swahili (English-only). Driven by the mamai-eval generator × prompt + safety evaluation.

Why

The mamai-eval matrix (Gemma 4 / 3n / Qwen × baseline/+G1/+G1+G2, end-to-end + faithfulness) plus a hand-adjudication of every flagged dangerous answer found:

• Gemma 3n is more helpful but materially less safe — genuine order-of-magnitude dosing/drug errors under these prompts (adjudicated as real, not judge artifacts).
• Gemma 4 + G1 fixes the deflection problem with ~zero dangerous answers — a net safety improvement over the previously-deployed 3n at comparable recall.

So Gemma 4 + G1 is the safest deployable on-device configuration.

Changes

Generator → Gemma 4 E4B

• app_config.json llm_model → gemma-4-E4B-it.litertlm
• intro_page.dart: download URL → ungated litert-community/gemma-4-E4B-it-litert-lm, SHA-256 pinned 0b2a8980…, size guard 4.92 → 3.66 GB, filename, startup item
• sync_models.sh, all "Gemma 3n" labels/comments → Gemma 4

Prompt → G1

• config/prompts/system_en.txt → the G1 prompt (deployed clinical system prompt)

Swahili removed (English-only)

• Deleted system_sw.txt, app_sw.arb, app_localizations_sw.dart
• Stripped sw code paths in RagPipeline.kt, removed _sw context labels + dead switcher strings, regenerated l10n English-only

Retriever (EmbeddingGemma) unchanged. Version 0.3.0+3 → 0.4.0+4. Docs updated.

Device test (real device 9da65e62) ✅

• flutter analyze: clean · release APK builds (203 MB)
• Intro shows "Gemma 4 on-device model — 3.66 GB", no language switcher
• In-app download from source succeeded + SHA-256 verified (marker == pinned 0b2a8980…) + model loaded
• Query "How do I measure fundal height?" → retrieved 3 guidelines (EmbeddingGemma) → grounded, cited, English clinical answer (TTFT 2.6 s), with G1's engaged + safety-escalation behavior

🤖 Generated with Claude Code