Faber 1.8.0

Install this version#

Pinned download for Faber 1.8.0. For the current release, use Start instead.

curl -fsSL -o faber.tgz \
  https://github.com/faberlang/releases/releases/download/faber-v1.8.0/faber-v1.8.0-aarch64-apple-darwin.tar.gz
tar -xzf faber.tgz
# The archive ships bin/ and share/; keep them together so the
# reader packs resolve beside the binary.
sudo mv bin/faber /usr/local/bin/faber
sudo mv share/faber /usr/local/share/faber
faber --version

Release notes#

Status: final

Minor product release on the odd-major 1.x development line (policy.md). Range: faber/v1.7.0 (5d7d42bf8) → pin 1eaf4ec68 (2026-08-15→2026-08-18). Headline: the compiler package surface moves into `radix-package` (RP1–RP5), packed quantized kernel plans land on Metal/NVVM (Q8_0 / Q4_K / Q5_K / Q5_0 / Q6_K / BF16 / F16 + QuantizedGather), shape-generics Phase 3+4 closes (unroll pass, guard-as-mask, SG-4B corpus, SG-4C archive), and the release pins the tested radix 0.83.0 companion.

Notes provenance: the release-lane charter says notes arrive pre-written. No v1.8.0 / v0.83.0 document was attached to handle 5dbc3c01. These notes are a factual reconstruction from git log faber/v1.7.0..1eaf4ec68 plus the carried 1.7.0 known-issues list. Disclose, do not invent.

Scale#

Reconstruct the full log:

git log faber/v1.7.0..1eaf4ec68 --oneline --no-merges

Major tracks#

radix-package (compiler package surface)#

Package compilation moves out of the faber product crate into radix-package: identities + API envelope (RP1), compiler manifest/lock view (RP2a), discovery + source loading (RP2b), import graph + library resolution + file interface (RP2c), multi-unit analysis (RP3), artifact planning + HIR emission (RP4), package MIR/FHIR/FMIR subset (RP5).

Packed quantized kernels (EXEC-02 / R-PACK)#

Frozen packed kernel plan/op ABI, FMIR wire-mirror, then Metal + NVVM emit for QuantizedMatMul (Q8_0, Q4_K, Q5_K, Q5_0, Q6_K, BF16, F16) and QuantizedGather. F16 admission + WGSL F16 elementwise IO.

Shape-generics closeout#

Named unroll pass for concrete-N for (3E), guard-as-mask differentiable subset (3F), SG-4B conversion-boundary + generic SDPA corpus, hir-rust ⇥ recovery, SG-4C goal closeout + archive at docs/archived/shape-generics/.

Product / tooling#

faber check fails when the emit def table misses a resolved def. Locale register for triga geometry layout helper. Library façade replaces the excluded stub. Canary smoke-test now proves codegen + faber run (not just packs) and the first-time-user journey.

Pin-window residuals (hand-65)#

Landed on the pin: rustfmt of five stage-1 compiler files, HIR compat measurement JSON regen, module-boundary SG-4B MIR recovery classified as a capability-gap, shape-generics archive.

Pin pair#

release-manifest.yaml pins this candidate against radix 0.83.0 (source 1eaf4ec68), hosts 0783406e2a93, cista 0.1.0 (no cista protocol call in this dispatch). publication.releaseTag is faber-v1.8.0. The --locked release build from these pins is the proof.

Radix companion (binary question)#

2026-08-25 update: the Radix binary release workflow has been removed. Faber is the only active binary release surface; historical Radix CLI archives can be backfilled manually if needed.

Live tree: radix is source+tag / library. There is no product radix binary. crates/radix has no [[bin]] (README: library façade; developer radix bin stays in radix-module until RTR4). cargo build -p radix --bin radix fails: "no bin target named radix in radix package".

radix/AGENTS.md still says verify with -p radix --bin radix — that step is stale. Live release.yml would also fail on the same -p radix --bin radix invocation. The remaining radix CLI is radix-module's developer binary, not a published component artifact.

This release therefore minor-bumps radix without a radix archive. Local tags v0.83.0 (AGENTS.md) and radix/v0.83.0 (workflow name) are created and not pushed. Pushing radix/v0.83.0 would fire a workflow that cannot build the named bin until that workflow is repointed. The shipped CLI is faber 1.8.0.

Known-issues (carried from 1.7.0; not re-derived)#

Operator ship_with_known_red (memo f12335e2) still applies. Carried verbatim:

• 97a33001 — GPU workload rung-0 floor regression: indexed-view ABI gap.
• f80df1a1 — swift emitter gap coverage: 96 fixtures classified, class table pending; post-release backlog.
• a828a988 — go emitter defects: union-typed error payloads (require) and nil-comparison codegen (sponte-vel).
• 8036dd57 — TS lane: codegen internal error on rewritten solum-lege-generic.fab (definition id 1000000 unresolved).
• b1f8eeba — rust e2e lane instans.fab cargo build produces no binary.
• 2624fea0 — channel bare-binary e2e harness host-dispatch gap (3 only/tempus fixtures, call/async golden trio).
• 54a93c8b — stage-4b new classified failures: forma idempotency x3, run-target merge-introduced, coreutils PARSE030, SEM004/SEM006 pinned-norma questions, hygiene x3.
• Environmental — RADIX_WASM_STUB_HOST: 16 wasm fixtures are environmental_skip because the run tier requires the retired radix-wasm-stub-host binary; supply via env when a prebuilt stub-host or external runner is available.

--full is recorded as a known-red-or-unrun receipt in the release closeout mail if the local tag-path suite is not completed before tag.

Verification#

Pin 1eaf4ec68 ran ./scripta/test --stage 1-2 (ok; clippy 0 errors) and cargo test -p radix-module-boundary (38 passed, 0 failed) before assembly. Post-build canary: assemble-dev-kit → package-archive → install-faber from the archive → installed faber --version + locale packs + faber check / faber run on the hello-world canary.

Known limitations#

Same as 1.7.0: cross-module string handles, no numeric-inference accuracy claim, Haskell/Python are source-emit surfaces, AMD stays fail-closed until ROCm clang is provisioned.

Deferred#

Onboarding Stages 4–8, numeric inference accuracy, AMD device-execution. Cista stays 0.1.0 (no protocol call).

---

All releases · Start