Add every hosted Gemma 4 model, and document adding your own without a release

[handstandsam]

Stacked on #228 — review that first. This PR's own changes are the two commits below; the base will collapse once #228 merges.

Summary

Every Gemma 4 model Google and Ollama serve is now built in. #225 added one; there are three, and you can pick any of them by id with no specs to look up:

Where	Ids
Gemini API (free tier)	gemma-4-12b-it, gemma-4-26b-a4b-it, gemma-4-31b-it
Ollama (local)	gemma4:12b, gemma4:26b, gemma4:31b

Google's own reference for the Gemini API ids: Gemma on the Gemini API.

And a new docs page says out loud that you never had to wait for a release to use a new model. A workspace trailblaze.yaml can declare any model its provider serves, and it takes effect on the next run — that was previously one sentence buried in the LLM reference. Adding a Model is now the page to hand someone who asks.

Two things worth flagging on #225

• gemma-4-31b-it and gemma-4-26b-a4b-it are both real, distinct models — 31B dense and 26B-A4B MoE. add gemma4 as a model for google #225's google.yaml declared the first and its hand-written LLM_MODELS.md row named the second, which reads like a typo but isn't. LLM_MODELS.md is generated from the provider yamls, so regenerating it (in Make waits settle instead of stalling, harden the iOS driver, and add the all: recording key #228) resolved the mismatch by silently dropping one real model. Both are back, plus the 12B.
• max_output_tokens is 32768, not the 65536 the Gemini models use. Google publishes no output cap for Gemma 4; 32768 is the largest figure any host documents.

The e2b / e4b edge builds are deliberately omitted from Ollama: 128K context and much weaker at the screenshot-driven tool calling the agent loop does. Worth knowing that gemma4:latest resolves to e4b, so name a size explicitly.

The docs change

Two files, one gap each.

New: docs/adding_a_model.md. How a workspace entry combines with the built-in list (matched by id — an unknown id adds a model, a known id overrides it field-by-field), the 131K/8K fallbacks you inherit by omitting the limits, what a model actually needs to be useful here (image input + tool calling), and recipes for hosted / local Ollama / your-own-gateway. Ends with how to contribute a model back to the built-in registry — including running :docs:generator:run, since LLM_MODELS.md is generated and hand-editing it is exactly what tripped up #225.

Rewritten opening of docs/configuration.md. The page titled "Configuration" documented Android instrumentation args and scripting callback timeouts and never mentioned trailblaze.yaml at all. It now opens with the file itself: where it lives and how it's discovered, that every section is optional, a table of the top-level keys (defaults, targets, toolsets, tools, providers, llm), the inline-or-ref: shape those lists accept, and one precedence table covering built-ins → user file → workspace file → persisted settings → env/flags. Existing sections are unchanged, below.

Cross-links added from index.md and llm_configuration.md.

Test plan

• ./gradlew :docs:generator:run — LLM_MODELS.md regenerates with all six rows, no other drift
• Doc anchors resolve against the pages they point at
• CI green

[handstandsam]

🤖 Stacked on top of this: #230 refreshes the whole built-in catalog to current OpenAI / Anthropic / Google models (and drops two ids that were already dead — gemini-3.1-flash-lite-preview is shut down, gemini-3.1-pro-preview-customtools isn't published by Google). It leans on the "no release needed to use a model" page added here as the migration path for anyone still on a dropped id.