chore(llm): swap Ollama model qwen3:8b → gemma4:e4b + bump container memory

[mrviduus]

Summary

Swap the local LLM model used for distractor / hint / metadata / tag generation from qwen3:8b to gemma4:e4b (Google Gemma 4 effective-4B, multimodal — text + vision + audio). Same ILlmService interface, no API changes.

What's in the PR

• 3 source defaults: Program.cs, OllamaLlmService.cs, VocabularyOptions.cs.
• 3 config files: Api/appsettings.json, Worker/appsettings.json, docker-compose.yml.
• 9 doc updates: README, CLAUDE.md, llm-provider-swap, 17-ai-auto-tags, PLAN-elevenreader-parity, plus stale gemma3:4b cleanup in docs/02-system/database.md, docs/01-architecture/README.md, docs/05-features/vocabulary-srs.md.
• TODO.md: drop "llama3, mistral, qwen, phi" speculation entry — choice is made.
• CHANGELOG.md: ### Changed entry under [Unreleased].

Container changes (this is the load-bearing part)

1. Pin ollama image to ollama/ollama:0.23.1. The floating latest tag was still serving 0.22.x in the registry, which does not recognise the gemma4 family — ollama pull gemma4:e4b returns "please download the latest version". 0.23.1 (released 2026-05-05) is the first stable that ships gemma4 support.
2. Memory limits raised: 4G/2G → 12G/8G. gemma4:e4b needs ~9.8 GiB RAM to load weights + KV cache (multimodal architecture is bigger than the prompt-author's claim suggested). Server has 31 GB total RAM so the headroom is fine.

Verification done locally

• dotnet build textstack.sln — green, 0 warn / 0 err.
• dotnet test tests/TextStack.UnitTests — 216/216.
• dotnet test tests/TextStack.Extraction.Tests — 169/169.
• dotnet test tests/TextStack.Search.Tests — 20/20.
• dotnet test tests/TextStack.IntegrationTests (vs running API) — 136/136.
• pnpm -C apps/web test — 419/419 (Vitest, 50 files).
• pnpm -C apps/web test:e2e --project=chromium tests/smoke.spec.ts — 5/5.
• apps/admin + apps/mobile tsc --noEmit — clean.
• Total: 965 tests passed.

Local Ollama JSON smoke could not run on this Mac (Docker VM has only 7.7 GiB; gemma4:e4b needs 9.8 GiB to load). Smoke will run on prod after merge — see post-deploy section below.

Post-merge / post-deploy actions (I will do these from SSH after CI deploy)

1. docker compose pull ollama — fetch the pinned 0.23.1.
2. docker compose up -d --force-recreate ollama — pick up the new image + memory limits.
3. docker compose exec ollama ollama pull gemma4:e4b — first-time pull (~10 min, 9.6 GB).
4. Smoke test: ollama run gemma4:e4b with the distractor prompt — verify it returns a clean JSON array (not markdown-fenced).
5. docker compose restart api worker — reconnect to fresh Ollama.
6. Save a vocabulary word via UI → confirm 5 distractors + hint + explanation appear in the DB.

Note: prod's Ollama volume is currently empty (0 models installed). qwen3:8b was set as the model name but never pulled — distractor generation has been silently falling back to random-word generation since deploy. This PR plus the post-deploy steps are also the moment we make LLM-backed features actually work in production.

Rollback

If gemma4:e4b produces broken output in prod:

ssh asus
cd ~/projects/onlinelib/textstack
echo 'Ollama__Model=qwen3:8b' >> .env
docker compose exec ollama ollama pull qwen3:8b   # ~5 min, 5 GB
docker compose restart api worker

Or revert this commit and redeploy.

🤖 Generated with Claude Code