Radix 0.77.0

Install this version#

No prebuilt archives were published for this version. It is listed here because its release notes are part of the record.

Release notes#

Minor release spanning 145 commits (v0.76.0..v0.77.0, 2026-07-20→2026-07-22). Driven by the MIR Swarm campaign: graphics pipeline reflection goes end-to-end, the AIR backward proof chain wires through the driver, WGSL workgroup reductions ship, and the workspace drops host/exempla siblings.

Scale#

Major tracks#

Graphics pipeline reflection#

End-to-end pipeline reflection from HIR shader-stage annotations through WGSL emission to JSON. The initial seam (MirGraphicsPipelineReflection + MirKernelShaderStage) was laid down, then extended from stub counts to full resource descriptors, multi-target color attachments, depth/stencil, stencil read/write masks, host-configurable primitive topology, fragment source extraction with combined vertex+fragment WGSL emission, and conflicting-stage diagnostics. Goal 01 completion (50d300b98) wires MirTrigaPipelineFact into the --reflection tool output.

• feat(mir): graphics pipeline reflection + draw prerequisites seam (140ec8707)
• feat(mir): assign MirKernelShaderStage during HIR→MIR lowering (961725982)
• feat(mir): surface resource counts in graphics pipeline reflection (d04c4bf81)
• feat(mir): add per-stage resource descriptors to graphics pipeline reflection (58662d138)
• feat(mir): multi-target color attachments in pipeline reflection (dbd6afbe2)
• feat(mir): depth/stencil state for graphics pipeline reflection (78c8c117c)
• feat(mir): stencil read/write masks and per-face stencil state (bd27ab6cf)
• feat(mir): host-configurable primitive topology builder (f792908d2)
• feat(mir): add emit_wgsl_fragment_entry_contract_with_targets to radix-mir-wgsl (6d8463878)
• feat(mir): implement fragment source extraction and combined vertex+ fragment WGSL emission (cc75ccc4a)
• feat(mir): emit diagnostic for conflicting shader-stage annotations (a159f5a10)
• feat(hir+mir): is_vertex/is_fragment HIR flags + Stage 4 lowering inventory (c6bc07db5)
• Goal 01 completion: pipeline-level reflection extraction from source (50d300b98)

AIR backward proof chain & autograd pipeline#

The generated-gradient pipeline wires through the driver for backward-annotated tensor units: generate_air_lane_backward runs the full proof chain (companion → AIR-to-MIR → result contract → eligibility → admission). Op-family dispatch routes elementwise, reduction, and matmul to the correct prove_* function. Fill/Vacuum type mismatches on the reduction path were repaired, and the MirBackwardResultContract now supports variable field counts.

• D-A-02 S1: wire AIR proof chain into driver for backward-annotated units (cada5a953)
• d-a-02 S2: source-level entry point + end-to-end exemplum (5f124b126)
• D-A-02 S3: elementwise op expansion — Div and Neg (66935a25f)
• D-A-02 S4: reduction op expansion — Sum + Mean (48b3e4502)
• D-A-02 S5: rank-2 matmul expansion (FINAL stage) (81568e64d)
• d-a-02 S4+S5 repair: wire dispatcher, fix Fill/Vacuum, add tests (e44733bc3)
• fix(radix): driver backward-skip repairs — Box::leak, scalar skip, detect_backward_op (c7479e1c2)
• fix(delivery): correct transpose oracle layout convention in D-A-08 (80e17f05b)
• delivery: D-A-07 domain unary backward (Exp + Log) — 4-unit serial graph (bb0580ecb)

Gradient ABI & LLVM backend#

ABI gradient handles hit both the surface (host-abi symbols + payload) and the LLVM backend emit for gradient calls. The GRADIENT_HANDLE_TYPE reserved constant was documented.

• D-A-03 U1: gradient ABI surface (host-abi symbols + ABI payload) (aef79e4d7)
• D-A-03 U2: LLVM backend emit for gradient calls (7175b728f)
• docs: document GRADIENT_HANDLE_TYPE as reserved constant (2494caa03)

WGSL workgroup reductions#

The MIR WGSL emitter now lowers Sum and Mean collection intrinsics to a workgroup reduction pattern: shared memory load, workgroup barrier, sequential tree reduce, and guarded output write (TensorMean divides by workgroup size). This promotion moves WGSL reductions from CheckedDebt to NativeSupport.

• feat(mir-wgsl): emit workgroup reduction for Sum/Mean Collection intrinsic (f4917c740)
• test: fix stage5 native blocker assertion after WGSL reduction promotion (51e6707d9)

MIR operations & infrastructure#

MirUnOp::Relu added with stepper dispatch (max(0, x) for Float/Int), dump label, validation, and exhaustiveness fixes. WASM backend gets a dedicated Relu type arm emitting f64.max. MirInputStepMode and Bgra8Unorm added to the radix-mir ABI types. MIR name analysis gains is_collection_op and is_collection_type helpers. The LM studio annotation exclusivity check was narrowed per CTO.

• radix: add MirUnOp::Relu with stepper dispatch, dump label, validation, and exhaustiveness fixes (961e04beb)
• fix(radix-mir-wasm): dedicated Relu type arm + negative tests (82c13ce93)
• mir: add MirInputStepMode + Bgra8Unorm to radix-mir ABI types (ea95b924d)
• fix(mir): annotation exclusivity + narrowed Stage 4 inventory per CTO (c6e49fd14)
• fix(mir): with_stencil fails closed when depth_stencil is None (a9e049429)

Spine design docs: device execution, reflection reciprocity, placement#

Three major design documents codify spine law for the GPU/host execution model:

• Placement execution contract (3fc85670b): 14-section spine law defining executable placement semantics — copy-in, readback, sync through __faber_gpu_v1_* ABI symbols, WebGPU first with LLVM parity follow-on.
• Reflection reciprocity (9a35b18b2): the compute↔graphics reflection schema contract — reciprocity matrix, shared root inventory, divergence rules, versioning policy, freeze process.
• Unified WebGPU host resource model (aead2be51): buffer identity via logicalId + generation, create-before-retire, queue-completion-gated destruction, compute-vs-render-pass ownership.

Delivery specs for each spine stage were also filed:

• delivery: D-SPINE-02 placement execution — 5-stage serial (8c011182b)
• delivery: D-SPINE-03 reflection reciprocity — 2-stage (00bb749bf)
• d-spine-01 S1: unified WebGPU host resource model design doc (aead2be51)
• spine: G-SPINE-03 S1 reflection reciprocity design doc (9a35b18b2)
• spine: placement execution contract design doc (G-SPINE-02 S1) (3fc85670b)
• fix(d-spine-02): correct abi_test.rs evidence path (f1ce397cf)

Workspace restructuring#

The crates/exempla e2e harness and the hosts/ directory (macos-arm64, webgpu-browser) were removed from the radix workspace. Platform hosts moved to the sibling faberlang/hosts repo; exempla moved to the sibling faber/crates/exempla. This unblocks standalone radix release builds that do not depend on the faber CLI sibling.

• refactor: move exempla e2e harness out of radix workspace (ef93b3e12)
• refactor: drop host products from radix workspace (59df81088)

Clippy pedantic deny (crates/radix)#

The main crates/radix package now denies pedantic clippy, matching the posture of the leaf crates. A multi-wave cleanup fixed ~hundreds of warnings across docs, must_use, map_or, imports, borrows, path borrows, and restore of rustdoc backtick code-links after a bad rewrite.

• chore(clippy): enable pedantic deny on crates/radix with structural allows (41fdff0bf)
• style(clippy): auto-fix mechanical pedantic on crates/radix (d6d5a67c6)
• style(clippy): must_use and remaining auto-fixable pedantic on radix (72e15b7fb)
• fix(clippy): parallel pedantic cleanup wave on crates/radix (734beec00)
• fix(clippy): second pedantic wave — docs, map_or, imports, borrows (af0575b91)
• fix(clippy): residual pedantic docs, allows, and path borrows (c3ae2faf3)
• fix(clippy): finish pedantic deny for crates/radix package (bb3954ecc)
• fix(clippy): clear residual non-pedantic warnings on crates/radix (dfa423465)

Codegen fixes (failable functions)#

Two related fixes ensure library-imported and locally-defined failable functions (⇥ / ∅) are correctly registered during Rust codegen. The extend_library_failable call was missing when a type table was present, and the HIR-level collect_failable_functions scan was gated behind a legacy branch. Both paths now run unconditionally, fixing roundtrip failures for norma::json, http_provider, and sqlite packages.

• fix(codegen): call extend_library_failable when type table is present (793908373)
• fix(codegen): always collect failable functions from HIR alongside library failable (58e83dc40)

MIR Swarm campaign planning#

47 docs(mir-swarm) commits cover Wave 0 retrospective, Wave 1 lock ledger rescore, P2 goal-check audits, P3 delivery audits, P1 supplementary goal forge, gap-forge queue management, and Wave 2 goal admission. Process docs codify the three-pass goal lowering pipeline, Vivi planning lessons, and fleet authority deferral. All 12 Wave 2 goals reached READY status.

Representative:

• docs(mir-swarm): add three-pass goal lowering pipeline (efbb3ae74)
• docs(mir-swarm): Wave 1 lock ledger rescore — 15 rows updated (2e0e27dd0)
• docs(mir-swarm): gap-forge-queue — 12 Wave 2 goals at READY (4a414189c)

Inference & fragment architecture#

The inference blocker matrix was updated to distinguish safetensors F32 (product-ready) from parked items (GGUF, tokenizer, quantization). A fragment architecture design doc (fragment-architecture.md) defines the fragment concept, rung ladder, tier model, and dtype policy (f32-only). Activations target rows were added to the tensor systems matrix.

• delivery: D-P-04 safetensors ingest — 4-stage serial unit graph (3ef24aa15)
• d-p-03 S1: fragment architecture design note (aef79e4d7)
• d-p-04 S4: update inference blocker matrix for safetensors availability (7175b728f)
• U4: Activations target rows + TensorOps blocker update (17473f18a)

Other changes#

• Fix backward parser: unknown_annotation_family now treated as skip, fixing 231+ test failures (2b8cb69a1)
• Ratchet hygiene budgets to match current source state (848564f00)
• Eliminate parallel cross-module FABER_LIBRARY_HOME env var race in tests (492add128)
• Fix str::replace patterns for 8-arg layout_matches (00909fda2)
• Restore varying validation + graphics pipeline reflection tests (fee457fc3)
• Add verifier-valid LLVM integration tests, remove dead code (fee457fc3)
• Fix WASM value_expr to return actual Wasm instruction type instead of MIR semantic type (834213a16)
• WGSL test: reject duplicate source vertex locations (d04c4bf81)
• Fix LLVM llvm.sqrt.f32/f64 intrinsic declaration before use (8e36c429f)
• Suppress false positive in diagnostic text ratchet for radix_lane_test (396497cb5)
• Fix doc_markdown backticks in tensor_operation_floor (6b526e62d)
• cargo fmt import reordering (4c0c8fc3b)
• Fix reflection-reciprocity doc accuracy + stale citations (707699089)
• Fix false counter invariant in design doc — destroyed is subset of retired (1a395bc24)
• release(radix): v0.77.0 — bump all 22 workspace crates to 0.77.0 (10e555055)

Verification#

Recorded on the release commit (bcdfd06fa):

Publish#

1. Push main (includes this notes file and version bump).
2. Annotated tag: git tag -a v0.77.0 -m "Radix v0.77.0"
3. Push tag: git push origin v0.77.0 (triggers .github/workflows/release.yml) or workflow_dispatch with tag v0.77.0.
4. Confirm faberlang/releases publishes radix-v0.77.0 multi-arch archives.

---

All releases · Install the current release