fix(qualification): accept semantic text task IDs in encoder parity

[jkzhang7]

Background

The shared STS accuracy qualification (embedding_vector_parity, benchmark
stsbenchmark_embedding_parity) chooses its Hugging Face reference mode and
benchmark operation from the candidate's task name, through a table in
_encoder_embedding_parity that only knows the retired encoding and
embedding names. A family migrated to the semantic Task SDK declares
text_to_pooled_features or text_to_embedding instead, so its qualification
fails before inference with:

QualificationError: embedding vector parity does not support candidate task 'text_to_pooled_features'

This was reported on #1584 (electra, mpnet and xlnet) by a reviewer, who
suggested a separate linked prerequisite so family migrations stay
self-contained. Every encoder family migrated to the semantic Tasks (including
#1584 and the other open encoder migrations) needs this routing.

Exit Criteria

• _encoder_embedding_parity accepts text_to_pooled_features and
  text_to_embedding and routes them exactly like the retired names: the
  first-token reference mode with encode, and the mean-pool reference mode
  with embed.
• The reference runner, gates and acceptance thresholds are unchanged.
• The retired names still work for families that have not migrated.
• Non-goal: any change to families, the benchmark worker, or the performance
  path (task_adapters.py already maps these Tasks).

Implementation

Affected component: shared qualification tooling only
(qualification_tests/benchmark_qualification/accuracy.py) and its test. No
family or runtime file is touched.

• accuracy.py: two entries added to the task table:
  text_to_pooled_features -> ("cls", "encode") and
  text_to_embedding -> ("embedding", "embed").
• tools/tests/test_model_benchmark.py: the existing parametrized
  test_encoder_accuracy_uses_task_semantics_without_model_specific_runner
  now also covers both semantic IDs, asserting the HF reference mode, the
  benchmark operation and the candidate requests.

The candidate side needs no change: for these Tasks the benchmark worker
already reports values, dim and feature_kind, which is what the
comparison reads.

Change categories

• CI or developer tooling

Validation

Commands and Results

$ PYTHONPATH=core/builder:apps/benchmark:. python3 -m pytest tools/tests/test_model_benchmark.py -q -k "encoder_accuracy or hf_encoder"
6 passed

$ PYTHONPATH=core/builder:apps/benchmark:. python3 -m pytest tools/tests/test_architecture.py -q
55 passed

Without the accuracy.py change, the two new test cases fail with the error
quoted above; with it they pass. Two unrelated tests in
test_model_benchmark.py (test_hf_translation_supports_replace_final_unknown_tokenizer)
fail on this machine on the unmodified tree because torch is not installed.

Real hardware: the three STS accuracy qualification cases that failed on
#1584, run through the repository's own runner on an RTX 4090 (SM 8.9, driver
595.91, image nvcr.io/nvidia/tensorrt:26.07-py3, TensorRT 11.1.0.106), with
this change merged onto the #1584 branch (426509c9):

python -m tools.model_benchmark run --model <model> --kind accuracy \
  --dataset stsbenchmark-test=<STSBenchmark/stsbenchmark_test.jsonl> \
  --runtime-root <family runtime> --trtmc-bench <trtmc-bench> --worker <trtmc_benchmark_worker>

The dataset file was fetched from mteb/stsbenchmark-sts and its SHA-256
(7c8927ee...e5bb) matches the digest pinned in
stsbenchmark_embedding_parity.yaml. All gates are the benchmark's own,
unchanged (100 samples, 50 pairs):

model                       status  vector_pass_rate  min_vector_cosine  max_pair_cosine_abs_delta
electra-base-discriminator  passed  1.0               0.999974           0.001622
all-mpnet-base-v2           passed  1.0               0.999981           0.002321
xlnet-base                  passed  1.0               0.999990           0.000486

HF vs candidate STS Spearman: electra -0.3244 / -0.3238, mpnet 0.9408 /
0.9404, xlnet -0.1678 / -0.1678 (the electra and xlnet base checkpoints are not
sentence-embedding models, so their absolute Spearman is low; the gate is
parity with the Hugging Face reference, not quality).

On the pod, pip install -e . failed, so trtmc-bench was invoked through a
two-line wrapper around python -m trtmc_benchmark.cli and the built trtmc
was linked at core/builder/tensorrt_model_connect/bin/trtmc; neither affects
the code under test.

Hardware, Environment, and Revisions

• Repository head: based on github/main at 5c675fb1.

Not Run / Remaining Gaps

• The performance qualification path is not changed and was not run here.

Contributor Self-Review

• I have completed a self-review of this change.

Notes For Future Readers

• Review order: accuracy.py (two table entries), then the test parameters.
• refactor(encoders): migrate electra, mpnet and xlnet to Task SDK #1584 depends on this for its STS accuracy cases; I will link them both.

Risk level

• Low

Two table entries in shared qualification tooling, covered by a test that fails
without them; the retired names are unchanged.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

_{Comment @coderabbitai help to get the list of available commands.}

[jkzhang7]

Folded into #1584 (commit d7ddd74) so that PR is correct on its own; closing this one.