v1.27.0_2

feat: drop low-fidelity q4/q4f16 release targets for jina nano (#27)

The 4-bit q4/q4f16 exports hold only ~0.62 pooled cosine vs fp32 (jina quantizes
BOTH the compute matmuls and the token embedding table to 4-bit; q4f16's fp16
compute also collapses individual samples to ~0.18). A last-token-pooling nano
model can't absorb that. Ship fp16 (~1.0) and the 8-bit `quantized` (~0.99) only.

- release.yaml: remove the q4 and q4f16 targets (now fp16 + quantized).
- test_ci_contracts: assert the 4-bit targets are gone (was: asserts q4f16 ships).

See issue #27 for the measurement (spike + structural inspect on branch
feat/gptq-4bit-quant).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Signed-off-by: Ruoshi Li <iceboundrock@gmail.com>

v1.27.0_1

fix: LD_PRELOAD shipped libonnxruntime.so so bench resolves its SONAME

The benchmark shard built and compiled fine but `bench` exited 127 at runtime:
the tarball ships only `libonnxruntime.so`, while the linked binary's DT_NEEDED
is the versioned SONAME (libonnxruntime.so.1.x.y), which is absent under -rpath,
so ld.so cannot resolve it. Preload the shipped .so when running bench; it
registers under its SONAME and satisfies the dependency.

Also surface bench stderr and exit non-zero on failure for clearer diagnostics,
and cast the AllocatorFree return in bench.c to silence a -Wunused-result warning.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Signed-off-by: Ruoshi Li <iceboundrock@gmail.com>

v1.0.16-c

v1.0.16-c

v1.0.16-b

fix: set quant to q4f16 to match prequantized model

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Ruoshi Li <iceboundrock@gmail.com>

v1.0.16-a

fix: use model_q4f16.onnx as base model instead of model_quantized.onnx

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Ruoshi Li <iceboundrock@gmail.com>

v1.0.15

fix: keep optimize pipeline for prequantized onnx

Signed-off-by: Ruoshi Li <iceboundrock@gmail.com>

v1.0.14

fix: use published quantized onnx for jina nano

Signed-off-by: Ruoshi Li <iceboundrock@gmail.com>

v1.0.13

docs: align optimization pipeline docs with current build

Signed-off-by: Ruoshi Li <iceboundrock@gmail.com>

v1.0.12

chore: use major GitHub Actions versions

Signed-off-by: Ruoshi Li <iceboundrock@gmail.com>

v1.0.10

fix: add .worktrees/ to .gitignore to exclude worktree files

Signed-off-by: Ruoshi Li <iceboundrock@gmail.com>