Release v0.7.0 · hyper-light/vorpal

v0.7.0 — the encoder becomes a candidate generator, on every platform

Always-on doc-side dense channel. Every index with the encoder enabled now embeds its referenced definitions at warm time — in the background, resumable across warms, never blocking a search — and fuses them as a fifth ranked list. The encoder stops being a reranker that can only reorder what the lexical channels found: Linux kernel encoder-enabled search 0.222 → 0.31–0.35 NDCG@10 (0.313 without the encoder), this repository 0.585 → 0.705. The fill's stop rule is structural (definitions something references; measured against the "or exported" variant and pinned on the data). --dense-budget-timeout 5m30s caps a fill round; <index>/dense.channel = off opts out.
Learned tier trains on a balanced kind set (every definition kind capped at the largest callable population, derived from the corpus): kernel learned tier 0.295 → 0.313, the first time it beats the lexical tier on the 8.9 M-node graph; CPython unchanged. Existing learned tiers retrain automatically (the policy label rides in the tier record).
vorpal tune measures the dense channel too and can pin a per-index verdict.

macOS: the doc-side path runs through Accelerate sgemm — 232 → 935–1,928 GFLOPS (AMX-bound, proven by shard sweep), embeddings identical to the reference lanes, bit-identical across thread counts.
x86-64 (Linux / Windows): owned AVX-512F and AVX2+FMA kernels, CPUID-detected — ~45× the previous portable path on CI's AVX2 runner (2 → 107–113 GFLOPS, bit-equal). AVX-512 compiles clean and awaits hardware for its first measurement.
GPUs (NVIDIA, AMD, Intel iGPU, Apple): one wgpu compute-shader GEMM over Metal / Vulkan / DX12 — no vendor SDK, ships inside the binary — with the weights resident on the device and the MLP block fused there; a GPU → CPU → reference ladder with typed fallback. On an M5 Max the fill runs 1.83× faster than Accelerate at 3.4× less CPU. Vulkan/DX12 are compile-verified and parity-tested on a software Vulkan device; hardware runs pending.
The query-side rerank stays on the fixed-order reference lanes on every platform — rankings are byte-identical to 0.6.1 wherever the dense channel is off.
int8 kernels (NEON sdot, AVX-VNNI, AVX-512-VNNI) ship but stay off: per-row activation quantization misses the retention bar on this encoder (recorded).

Walk reuse for incremental saves of giant files (0.6.1's headline) is unchanged.
MSRV is now honestly 1.98 (the tree already relied on 1.88 features); three newer style lints unlocked by the bump are allowed at the workspace level pending hygiene work.
docs/wip/ENCODER_RESEARCH.md: the primary-source survey behind this release, including the measured-and-rejected options (LateOn-Code multi-vector channel, cross-encoder reranking, static distillation, richer candidate surfaces).

Full Changelog: v0.6.1...v0.7.0