Release pipeline (#741)

Paused, 19 October 2026: .github/workflows/release.yml skips all three jobs. The legacy-crate repair restored the compliant library surfaces of the graph and root packages, but neither release asset can be built under the D23/D25/D29/D30 bans. The r4 CLI serves R4G1 through the retired transformerless compiler/runtime and the source-backed teacher, and the wasm frontend is the R4G1 browser engine D36 already paused. The commands below (r4 package-release-bundle, r4 install-release) are therefore not built. The M8 rewrite for a D11 served artifact (UORLUT01) re-enables the pipeline.

Preserved runtime reference: this pipeline packages the existing TLA/R4G1 frontends and bundle contracts. It is not the release pipeline for a route-native geometric-intelligence product, and a successful legacy release does not establish source-free attention, generation, correctness, or chat. Current product sequencing is defined by the Geometric Intelligence Programme.

Adopted by maintainer decision on 2026-08-19 (recorded on #741/#655): the "package + build release pipeline now" option — both frontends, a GitHub Release as the distribution artifact, an explicit verified fetch command, and a versioning convention that binds each tag to its exact code and bundle identity. This pipeline publishes nothing by itself: every piece below is inert until the maintainer pushes a v* tag, and even then the release starts as a draft that only the maintainer publishes.

Versioning convention

A release tag vX.Y binds, in one auditable place:

code identity — the tag's commit SHA (in the release notes and on the tag itself);
contract identity — the inference-contract version from docs/inference_contract.md (in the release notes);
bundle identity — the blake3 component digests (normative graph, sections-absent and label-shuffled control graphs, signature artifact, plain-TLA comparator store, tokenizer, score/compile reports, normative deployed-quality report, raw cross-surface parity evidence, and raw witness replay) declared by the attached schema-2 release-bundle.json. The quality report in turn content-binds the corpus pair, tokenizer adapter, compiler revision/configuration, selector semantics, held-out partition, controls, raw parity/replay evidence, and exact production decode profile. Schema-1 manifests remain readable only as historical research evidence and cannot authorize serving.

A tag with no bundle assets attached is a code-only release; the verified fetch command simply reports the missing asset.

What the tag triggers (`.github/workflows/release.yml`)

Pushing vX.Y creates a draft release and attaches both frontends:

r4-vX.Y-x86_64-unknown-linux-gnu.tar.gz — native CLI, linux;
r4-vX.Y-aarch64-apple-darwin.tar.gz — native CLI, macOS arm64;
r4-wasm-vX.Y.tar.gz — the wasm frontend (wasm-pack build, the same build the Pages deploy runs);
a .sha256 beside each.

CI cannot compile the model (hours of teacher-bound compute against a pinned snapshot), so the model bundle is attached by the maintainer:

Maintainer steps to cut a release

Pick the bundle (current candidate per the decision record: the smollm2-360m-broad compile) and package its sidecar:

r4 package-release-bundle \
  --compiled .uor-models/compiled/<name> \
  --model-id r4 --capability instruction-chat \
  --source .uor-models/sources/<source> \
  --tokenizer-family hf-byte-bpe --tokenizer-version 1 \
  --compiler-revision "$(git rev-parse HEAD)"

Archive exactly the attested component files (the packaged layout — nothing else; r4 install-release refuses archives carrying any unattested entry). Local invocation journals, resource sidecars, and terminal transcripts are operational evidence and are deliberately not packaged:

cd .uor-models/compiled/<name>
tar -czf /tmp/release-bundle.tar.gz \
  graph/score.r4g1 graph/score_sections_absent.r4g1 \
  graph/score_label_shuffled.r4g1 tless_artifacts.bin tless_store.bin \
  tokenizer.bin \
  graph/score_report.json graph/deployed_quality_report.json \
  graph/cross_surface_parity.json graph/witness_replay.json \
  graph-cover/cover_report.json corpus.meta corpus.records \
  tokenizer_adapter.json

Tag and push: git tag vX.Y && git push origin vX.Y — the workflow drafts the release and attaches the frontend builds.

Attach the bundle and publish:

gh release upload vX.Y /tmp/release-bundle.tar.gz \
  .uor-models/compiled/<name>/release-bundle.json
gh release edit vX.Y --draft=false

The explicit verified fetch

r4 install-release --tag vX.Y            # from UOR-Foundation/uor-r4
r4 install-release --tag vX.Y --repo owner/name --name custom-install

install-release (src/release_install.rs) downloads the two bundle assets, requires exactly the schema-2 production inventory above, hard-verifies every manifest component digest, and independently reproduces the deployed-quality bindings from the extracted graph, teacher artifact, exact plain-TLA comparator store, corpus, tokenizer, adapter, reports, both planted control graphs, raw cross-surface rows, and independently replayed witness rows. It refuses unknown files, symlinks, missing/tampered raw evidence, legacy/sample/off-serving evidence, or any identity mismatch, never overwrites an existing install, and only then moves the bundle into .uor-models/compiled/<name>. A failure at any step installs nothing. The fetch is always explicit: no serving path, first request, or setup step triggers it — exactly #655's "first request must not silently download" scope line.

What this pipeline deliberately does not do

It does not flip any serving default or canonical model identity — that is #655-F, separately gated on its own preconditions and a fresh maintainer sign-off.
It does not auto-publish: drafts require the maintainer.
It does not compile models in CI.

UOR-R4