fix(native): improve backend diagnostics and context handling

[justinchuby]

Summary

• make native backend failures identify unsupported components and operators
• preserve model context limits in pipeline loading
• improve native/ORT support-library diagnostics

Validation

• pipeline executor test suite

[justinchuby]

Rebased again onto latest main after its sibling PRs landed and re-resolved the four conflict hunks in pipeline/mod.rs. All four are independent additions -- generation_defaults (from #1162) and max_sequence_length (this PR) touching the same struct field list, accessor block, �uild locals, and constructor -- so the resolution is to keep both on every hunk. The accessor hunk needed a closing brace inserted, since the two sides had been sharing one.

Re-validated after the rebase: cargo test -p onnx-genai-engine --no-default-features --features native-backend --test pipeline_executor -> 2 passed, including pipeline_exposes_declared_context_limit. The reviewer had already falsified that test (dropping the .or(self.max_sequence_length) reddens it with left: None, right: Some(16)).

Merging.

[justinchuby]

Closing in favour of #1200, which carries this exact commit on a branch based directly on main.

This PR could not be merged: it is the base of #1161, and GitHub blocks both merge and base changes for stacked pull requests, pointing at an asynchronous merge endpoint this repository returns 404 for. The content was reviewed and validated, so it is not being dropped -- only re-landed off the stack, which also unblocks #1161.

[codecov]

Codecov Report

❌ Patch coverage is 0% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 79.85%. Comparing base (4435c51) to head (b2f295b).
⚠️ Report is 20 commits behind head on main.

Files with missing lines	Patch %	Lines
...es/onnx-runtime-ep-cpu/src/kernels/matmul_nbits.rs	0.00%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1158      +/-   ##
==========================================
+ Coverage   79.60%   79.85%   +0.25%     
==========================================
  Files         357      359       +2     
  Lines      153790   157326    +3536     
  Branches   153790   157326    +3536     
==========================================
+ Hits       122419   125631    +3212     
- Misses      26850    27135     +285     
- Partials     4521     4560      +39     

Flag	Coverage Δ
mlas	85.39% <ø> (?)
offline	79.74% <0.00%> (+0.14%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
...es/onnx-runtime-ep-cpu/src/kernels/matmul_nbits.rs	75.88% <0.00%> (-0.14%)	⬇️

... and 18 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[github-actions]

🔴 Benchmark Regression Detected

Comparison of criterion micro-benchmarks: PR head vs merge-base, measured on the same runner in the same job (base first → PR second).

ℹ️ Absolute times are informational only — they vary with runner load. The % change column is the reliable signal because both sides ran under identical conditions.

Status	Scenario	Base	PR	Change
🔴	matmul/small_generic_f32_threads=8/1x256x256	45.10 µs	90.50 µs	+100.7%
🔴	matmul/small_generic_bf16_threads=8/1x256x256	37.07 µs	51.27 µs	+38.3%
🔴	matmul/small_generic_f16_threads=8/1x256x256	36.15 µs	48.76 µs	+34.9%
⚠️	grammar_masking/llguidance_compute_mask/32	70.30 µs	88.00 µs	+25.2%
✅	reduce_mean/medium_f32_threads=1-internal/65536	235.67 µs	269.02 µs	+14.2%
✅	kv_cache/alloc_dealloc_pages	37.39 µs	39.78 µs	+6.4%
✅	logit_processing/seven_processor_chain_per_step	299.82 µs	318.20 µs	+6.1%
✅	block_quantized_matmul_cached_dense/mxfp4_cached_dense_repeated_call/1x1024x1024	42.43 µs	44.94 µs	+5.9%
✅	add/large_bf16_threads=1-internal/4194304	1.54 ms	1.62 ms	+5.2%
✅	qwen3_sampling_processors/top_k_top_p_fast	612.11 µs	643.76 µs	+5.2%
✅	matmul/medium_generic_f32_threads=1/32x512x512	2.25 ms	2.36 ms	+5.0%
✅	gather/small_f32_threads=1-internal/4096	622.8 ns	653.0 ns	+4.9%
✅	matmul/large_generic_bf16_threads=8/32x1024x1024	1.25 ms	1.31 ms	+4.8%
✅	add/small_bf16_threads=1-internal/1024	411.3 ns	424.9 ns	+3.3%
✅	add/small_f16_threads=1-internal/1024	424.3 ns	438.0 ns	+3.2%
✅	add/medium_f32_threads=1-internal/262144	22.62 µs	23.18 µs	+2.5%
✅	reduce_mean/large_f32_threads=1-internal/262144	922.63 µs	942.90 µs	+2.2%
✅	matmul/large_generic_f32_threads=8/32x1024x1024	3.58 ms	3.65 ms	+1.9%
✅	gather/medium_bf16_threads=1-internal/32768	2.42 µs	2.46 µs	+1.8%
✅	block_quantized_matmul_cached_dense/mxfp4_uncached_dequant_each_call/1x1024x1024	507.04 µs	514.86 µs	+1.5%
✅	matmul/medium_generic_f16_threads=1/32x512x512	32.67 µs	32.85 µs	+0.5%
✅	qwen3_sampling_processors/top_p_fast_after_top_k	534.24 µs	536.28 µs	+0.4%
✅	reduce_mean/small_f32_threads=1-internal/4096	13.99 µs	14.02 µs	+0.2%
✅	block_quantized_moe_cached_dense/mxfp4_uncached_expert_dequant_each_call/rows=1,H=256,I=256,E=4,top_k=1	431.97 µs	431.49 µs	-0.1%
✅	add/large_f16_threads=1-internal/4194304	1.54 ms	1.53 ms	-0.4%
✅	matmul/large_generic_bf16_threads=1/32x1024x1024	1.81 ms	1.80 ms	-0.6%
✅	matmul/large_generic_f16_threads=8/32x1024x1024	78.39 µs	77.51 µs	-1.1%
✅	add/medium_bf16_threads=1-internal/262144	96.22 µs	94.85 µs	-1.4%
✅	sampling_latency/min_p_per_token	198.94 µs	195.68 µs	-1.6%
✅	matmul/large_generic_f16_threads=1/32x1024x1024	74.48 µs	72.77 µs	-2.3%
✅	gather/small_f16_threads=1-internal/4096	463.2 ns	448.3 ns	-3.2%
✅	sampling_latency/top_p_per_token	370.49 µs	354.86 µs	-4.2%
✅	gather/large_f32_threads=1-internal/131072	30.93 µs	29.53 µs	-4.5%
✅	add/small_f32_threads=1-internal/1024	187.5 ns	178.7 ns	-4.7%
✅	qwen3_sampling_processors/top_k_top_p_full_sort_baseline	5.88 ms	5.55 ms	-5.6%
✅	matmul/large_generic_f32_threads=1/32x1024x1024	9.15 ms	8.55 ms	-6.6%
✅	qwen3_sampling_processors/top_k_full_sort_baseline	2.17 ms	1.98 ms	-8.7%
✅	matmul/medium_generic_bf16_threads=1/32x512x512	535.70 µs	489.00 µs	-8.7%
✅	qwen3_sampling_processors/top_p_full_sort_after_top_k_baseline	3.59 ms	3.26 ms	-9.1%
✅	gather/small_bf16_threads=1-internal/4096	501.1 ns	454.7 ns	-9.3%
✅	sampling_latency/greedy_per_token	3.40 µs	3.08 µs	-9.5%
✅	tokenization/decode_tokens_per_second	6.48 ms	5.84 ms	-9.8%
✅	add/large_f32_threads=1-internal/4194304	623.88 µs	560.06 µs	-10.2%
✅	matmul/small_generic_f16_threads=1/1x256x256	33.24 µs	29.66 µs	-10.8%
✅	block_quantized_moe_cached_dense/mxfp4_cached_dense_expert_repeated_call/rows=1,H=256,I=256,E=4,top_k=1	176.60 µs	156.52 µs	-11.4%
✅	gather/medium_f16_threads=1-internal/32768	2.58 µs	2.28 µs	-11.4%
✅	sampling_latency/top_k_per_token	55.47 µs	48.43 µs	-12.7%
✅	matmul/small_generic_bf16_threads=1/1x256x256	34.96 µs	30.48 µs	-12.8%
✅	qwen3_sampling_processors/top_k_partial_selection	158.11 µs	135.82 µs	-14.1%
✅	block_quantized_matmul_cached_dense/mxfp4_preexpanded_dense_oncelock_like_proxy/1x1024x1024	47.94 µs	40.90 µs	-14.7%
✅	add/medium_f16_threads=1-internal/262144	116.36 µs	99.04 µs	-14.9%
🟢	matmul/medium_generic_f16_threads=8/32x512x512	38.34 µs	31.82 µs	-17.0%
🟢	gather/large_f16_threads=1-internal/131072	16.69 µs	13.78 µs	-17.4%
🟢	tokenization/encode_tokens_per_second	421.65 µs	347.89 µs	-17.5%
🟢	gather/medium_f32_threads=1-internal/32768	4.40 µs	3.61 µs	-18.1%
🟢	matmul/medium_generic_f32_threads=8/32x512x512	1.26 ms	996.06 µs	-20.8%
🟢	matmul/medium_generic_bf16_threads=8/32x512x512	460.44 µs	353.99 µs	-23.1%
🟢	gather/large_bf16_threads=1-internal/131072	18.55 µs	14.22 µs	-23.3%
🟢	matmul/small_generic_f32_threads=1/1x256x256	51.97 µs	35.22 µs	-32.2%

Visual flags: ⚠️ ≥ 15% slower, 🔴 ≥ 30% slower — calibrated against measured runner noise (~27% worst-case on multi-threaded matmul)

Host info

CPU: Apple M1 (Virtual)
Cores: 3
OS: Darwin 25.5.0 arm64
Rust: rustc 1.97.1 (8bab26f4f 2026-07-14)
Load avg: { 4.83 3.97 4.81 }

What this cannot catch

• Regressions in code paths not covered by these benchmarks (e.g., end-to-end decode with a real model)
• Sub-threshold regressions that compound over multiple PRs
• Performance changes that only manifest under GPU execution
• Latency changes in the ORT integration path (these benchmarks exercise the native Rust kernels)