recipe(lt-wikidata-comp-zh): add feature extraction recipes

[kujin66]

Adds recipe coverage for dell-research-harvard/lt-wikidata-comp-zh as a BERT feature-extraction model. The current origin/main baseline already exports the checkpoint, so this is an L0 recipe-only contribution with measurable L1 value from the checked-in w8a16 quantized recipe running successfully on CPU.

1. Recipe path(s)

• examples/recipes/dell-research-harvard_lt-wikidata-comp-zh/feature-extraction_fp16_config.json
• examples/recipes/dell-research-harvard_lt-wikidata-comp-zh/feature-extraction_w8a16_config.json

2. README row

• Added dell-research-harvard/lt-wikidata-comp-zh | feature-extraction to examples/recipes/README.md.
• Updated total from 75 to 76.

3. Build output dir

• Baseline: temp/baseline_lt_wikidata_comp_zh
• Recipe fp16: temp/lt_wikidata_comp_zh_fp16_recipe_fixed
• Recipe w8a16: temp/lt_wikidata_comp_zh_w8a16_recipe
• L1 perf JSON: temp/lt_wikidata_comp_zh_w8a16_perf_cpu.json

4. Build log

• Baseline command: uv run winml build -m dell-research-harvard/lt-wikidata-comp-zh -o temp/baseline_lt_wikidata_comp_zh --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild
• Baseline result: Build complete in 59.8s; final artifact temp/baseline_lt_wikidata_comp_zh/model.onnx.
• Recipe fp16 command: uv run winml build -c examples/recipes/dell-research-harvard_lt-wikidata-comp-zh/feature-extraction_fp16_config.json -m dell-research-harvard/lt-wikidata-comp-zh -o temp/lt_wikidata_comp_zh_fp16_recipe_fixed --no-analyze --no-compile --rebuild
• Recipe fp16 result: Build complete in 53.6s; FP16 stage ran in 7.7s, artifact 194.1 MB, 202 FLOAT16 initializers, I/O types preserved.
• Recipe w8a16 command: uv run winml build -c examples/recipes/dell-research-harvard_lt-wikidata-comp-zh/feature-extraction_w8a16_config.json -m dell-research-harvard/lt-wikidata-comp-zh -o temp/lt_wikidata_comp_zh_w8a16_recipe --no-analyze --no-compile --rebuild
• Recipe w8a16 result: Build complete in 81.9s; quantized artifact 113.6 MB.

5. Appended findings

• No in-repo model_knowledge file is included in this Lane B model PR. Model-specific finding to record externally: bert: dell-research-harvard/lt-wikidata-comp-zh resolves as BertModel/feature-extraction, not tabular-classification; fp16 and w8a16 CPU recipes build, and w8a16 runs perf.

6. Optimum-coverage probe

{
  model_type: bert,
  architectures: [BertModel],
  vendor: [],
  after_winml: [feature-extraction, fill-mask, multiple-choice, question-answering, text-classification, token-classification],
  added_by_winml: [feature-extraction, fill-mask, multiple-choice, question-answering, text-classification, token-classification]
}

7. Claimed (Effort, Goal, Outcome)

• Effort: L0 recipe-only, using existing WinML BERT registration.
• Goal ceiling: L1, because baseline already covers L0 and the recipe delta is quantized w8a16 CPU perf evidence.
• Outcome: L0, recipe JSON plus README row only.
• Target EPs: cpu.
• Baseline freshness: origin/main and baseline HEAD were both 904c20109ab31d737c5132c3686ab404236edd57.
• PR head after reviewer fix: 937b000e.
• winml --version: 0.2.0.

8. Goal-ladder verdict table
   | Tier | Verdict | Evidence |
   |---|---|---|
   | L0 | PASS | WinMLBuildConfig.from_dict(...).validate() passed for both recipes; fp16 recipe build completed in 53.6s and emitted 202 FLOAT16 initializers; w8a16 recipe build completed in 81.9s; ONNX input shapes are input_ids, attention_mask, token_type_ids all [1,512]; output is last_hidden_state [1,512,768]. |
   | L1 | PASS | CPU EP perf on temp/lt_wikidata_comp_zh_w8a16_recipe/model.onnx completed with CPUExecutionProvider; providers snapshot was [DmlExecutionProvider, CPUExecutionProvider]; mean latency 595.581 ms, p50 654.255 ms, throughput 1.68 samples/sec, RSS total delta 285.93 MB; perf JSON saved to temp/lt_wikidata_comp_zh_w8a16_perf_cpu.json. |

9. Methodology-evolution declaration

• No new methodology friction observed for this model PR. A reviewer-style pass caught the fp16 recipe needed an explicit mode: fp16; this is already covered by existing reviewer rule _meta-014, so no new _meta-NNN or skill-file edit is required.
• No skill files are edited in this Lane B model PR.

[kujin66]

Reviewer-style verification (not a formal GitHub approval; same account authored the PR):

• PR branch head: 937b000e891240234ec0ecfb0856e3501b486630.
• Scope check: diff is limited to examples/recipes/README.md plus the two examples/recipes/dell-research-harvard_lt-wikidata-comp-zh/*.json recipe files. No source, tests, or skill files are included.
• Baseline freshness: PR body cites origin/main/baseline HEAD 904c20109ab31d737c5132c3686ab404236edd57; current origin/main was refreshed and matches that commit.
• Recipe check: feature-extraction_fp16_config.json now uses quant.mode=fp16; rebuilt artifact temp/lt_wikidata_comp_zh_fp16_recipe_fixed/model.onnx has 202 FLOAT16 initializers and preserved I/O. feature-extraction_w8a16_config.json artifact contains QDQ nodes (QuantizeLinear=428, DequantizeLinear=630).
• L0/L1 evidence: PR body includes build output dirs, build-complete lines, structural I/O shapes, CPU provider snapshot, and w8a16 perf numbers.
• CI: all checks are passing at time of comment: license/cla, lint, Analyze (Python), CodeQL, and all test jobs.

Remaining gate: external maintainer/reviewer approval and conversion from Draft to Ready.