Bundle OpenCodeReview v1.12.12 (accept unlisted models)

[testudor]

Why

pruefbyte 0.1.0 bundles OpenCodeReview (ocr) v1.12.10. That version rejects any --model that is missing from a built-in provider's model list. A CI review configured with provider: openai-responses, model: gpt-6.1-sol fails with:

Error: resolve LLM endpoint: resolve OCR config file: model "gpt-6.1-sol" is not available for provider "openai-responses"; available models: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna

What changed in ocr (v1.12.10 → v1.12.12)

The diff is 6 commits (compare):

• v1.12.12, fix(llm): allow unlisted OpenRouter model overrides (#1610)
  • In internal/llm/resolver.go, a --model that is not listed for a built-in provider no longer fails. ocr now prints [ocr] WARNING: model "…" is not in the suggested models for provider "…"; the provider will validate it and sends the request.
  • Custom providers keep the old check.
  • The openai-responses list itself is unchanged.
• v1.12.11, feat(allowlist): add Jinja template support (#1056)
  • .jinja2 and .j2 files are now reviewed, with a new built-in rule (jinja.md).
  • Repositories with Jinja templates will see findings on them.
• v1.12.11, other changes: they don't affect pruefbyte.
  • IDE extensions: provider catalogs generated from the Go registry (#1212), merge-commit file listing (#1544)
  • viewer pagination (#1576)
  • Windows make output name (#1573)
  • docs

Checked against what pruefbyte relies on:

• Asset names and sha256sum.txt format: unchanged. The Dockerfile and go run ./internal/ocrbin/fetch work as before.
• Flags and config keys pruefbyte passes: unchanged. No ocr CLI code changed; --provider/--model keep their meaning.
• JSON output and rule-file format: unchanged. No review or output code changed. The rule schema is the same, and the only addition is the Jinja system rule.
• Minimum git version: still 2.41.
• Provider registry: identical in both versions (29 presets). pruefbyte's builtinProviders already matched it; only the version in its comment changes.

Changes in this PR

• internal/ocrbin/VERSION: v1.12.10 → v1.12.12. This file drives both the Docker image and the release binaries.
• Version references in the comments of internal/config/config.go and internal/ocrbin/ocrbin.go.
• README.md and pruefbyte.example.yml: llm.model can be any ID the provider serves; ocr's built-in lists are only suggestions.

pruefbyte itself never validated the model against a list; it only requires llm.model to be set. No behaviour change beyond the ocr bump.

Verification

• Static checks and tests on a clean LF clone in Linux containers:
  • gofmt -l . prints nothing.
  • golangci-lint run ./... (v2.14.0): 0 issues.
  • go test -race ./...: all packages pass.
  • go vet passes with and without -tags embedocr.
• Fetch: go run ./internal/ocrbin/fetch downloaded all six v1.12.12 binaries and verified each against the release sha256sum.txt.
• Release build: goreleaser release --snapshot --clean produced all six archives. The Windows and Linux snapshot binaries both report ocr v1.12.12 (built in).
• Model check, before and after, in a scratch repo whose target branch has a .pruefbyte.yml with provider: openai-responses, model: gpt-6.1-sol, running pruefbyte local --target main with a dummy key:
  • Released 0.1.0 (ocr v1.12.10) fails with model "gpt-6.1-sol" is not available for provider "openai-responses", reproducing the report.
  • This branch (ocr v1.12.12) prints [ocr] WARNING: model "gpt-6.1-sol" is not in the suggested models for provider "openai-responses"; the provider will validate it, gets past "resolve LLM endpoint", and fails at the provider.
  • ocr llm test with the same config shows the provider's answer: POST https://api.openai.com/v1/responses: 401 Unauthorized … invalid_api_key.
• Docker: docker buildx build --platform linux/amd64,linux/arm64 succeeds. Each architecture, loaded and run (arm64 under QEMU), reports open-code-review v1.12.12 and ships both license texts.

Not verified

• A real review with gpt-6.1-sol: no OpenAI key was available in the environment, so I couldn't confirm the model ID works at OpenAI or that the review output is sound.
• macOS binaries: built, but not run.

Release (maintainer, after merge)

Tagging is not part of this PR. After merging, push v0.1.1 on the merge commit. The tag's CI run produces:

• GitHub release v0.1.1: pruefbyte_{linux,darwin}_{amd64,arm64}.tar.gz, pruefbyte_windows_{amd64,arm64}.zip and checksums.txt, all with ocr v1.12.12 embedded.
• Image tags: ghcr.io/feinarbyte/pruefbyte:0.1.1, :0.1 (moves from 0.1.0), and :sha-<commit>, each for linux/amd64 and linux/arm64.

:latest follows main, so it gets ocr v1.12.12 as soon as this PR is merged, before the tag.

🤖 Generated with Claude Code