Consolidate reusable prefab, model-adapter and checkpoint improvements

## Category consolidation

Requested by Joel Z. Leibo in [issue #373](https://github.com/google-deepmind/concordia/issues/373#issuecomment-5621250114). This is the single continuing issue for this category. Implementation and subsequent improvements stay in [PR #356](https://github.com/google-deepmind/concordia/pull/356). Approved PRs #297, #302, #304 and #332 are exempt and remain unchanged.

## Scope

- Publish complete checkpoint JSON through same-directory atomic replacement, preserving a previous file on serialization or I/O failure.
- Share the existing optional-context assembly between basic and minimal prefabs, and retain the normal Concat name-prefix option.
- Add opt-in timeout/output bounds and JSON response schemas to the existing Ollama adapter. Default behavior remains unchanged.

These are existing-library improvements extracted from the Bellwether stack, not game-specific mechanics. The human acting-policy prerequisites remain separately approved in PRs #302 and #304. Approved checkpoint-input ownership #332 and viewer #297 remain untouched.

## Consolidated review history

No additional standalone PR in this category; prerequisite pieces are extracted from the former Bellwether stack.

Superseded issues and PRs retain their original descriptions and point here. Consolidation closes duplicate review threads, not claims that pending code has landed upstream.

Start by reviewing PR #356 and the consolidated scope in this issue; no source files or tests are named. Compare the pending implementation with the checkpoint, prefab, and Ollama adapter requirements, and consider the work done only when those listed improvements are addressed without changing the exempt PRs and ownership described here.

Written by the indexing model from the issue text.