PR: Helsinki-NLP/opus-mt-en-ru — translation recipe pair (fp32, CPU) — Goal-L2-encoder closed

Iter: 6 (composite recipe pair shipped iter-5 as marian-003; this PR adds the Goal-L2-encoder + L1-CPU evidence on top)
Producer: main agent (2026-06-23)
Claimed tier: (Effort = L0★, Goal = L2-encoder, Outcome = L0)

Summary

This PR ships the Helsinki-NLP/opus-mt-en-ru translation recipe pair (encoder + decoder). It is the FIRST seq2seq composite pair contributed to the recipe catalog, and the first Marian-family entry. The recipe was generated via winml config --task translation (per _meta-020 composite-expansion gate); both halves build cleanly on CPU at fp32. Goal-L1-CPU PASSes on both halves; Goal-L2 cosine = 1.000000 on the encoder (PT-vs-ONNX). Goal-L2 on the decoder is DEFERRED-HARNESS per _meta-018 — see verdict table. No source-code changes.

Per _meta-020, encoder + decoder ship as ONE PR with a per-half verdict matrix.

1. Recipe files

• examples/recipes/Helsinki-NLP_opus-mt-en-ru/translation_fp16_encoder_config.json
• examples/recipes/Helsinki-NLP_opus-mt-en-ru/translation_fp16_decoder_config.json

Note on filename: fp16_* is cosmetic per _meta-014 — quant: null means fp32 weights ship. winml perf correctly reports Model Precision: fp32 (see L1-CPU evidence below). The cosmetic filename is retained for catalog consistency.

2. README index row

examples/recipes/README.md — row to add for Helsinki-NLP/opus-mt-en-ru | translation | composite (encoder + decoder) | recipe pair.

3. Build output directory + artifact inventory

temp/marian_build/{encoder,decoder}/ (gitignored — referenced by path for reviewer re-execution):

External-data layout check (_meta-023): both halves under 2GB ProtoBuf limit ⇒ inline weights, no .data shard. N/A — vacuous PASS.

Encoder/decoder cross-attention alias check (_meta-025): encoder output = encoder_hidden_states (shape [1,512,512]); decoder input encoder_hidden_states (shape [1,512,512]). Direct name + shape match. PASS.

4. Build log

Build logs at temp/marian_build/{encoder,decoder}/build.log (per marian-003 mechanism_notes). Iter-6 reused iter-5 artifacts unchanged — recipe is byte-identical to the marian-003 commit; no re-build needed.

5. Appended findings

Per-model — model_knowledge/marian.json

• marian-003 — VALIDATED L0★ build closure (iter-5).
• marian-005 — VALIDATED Goal-L1-CPU + Goal-L2-encoder cosine = 1.0 (this PR's primary evidence).
• marian-006 — PR-mining cross-references (composite gate _meta-020, encoder alias _meta-025, external-data _meta-023, --ep-options retry _meta-026, task-consistency _meta-028).

Skill-meta — skill_meta/findings.json

This PR does not introduce new _meta-NNN findings. The iter-6 methodology evolution (_meta-019..037) ships separately on the skills branch (Lane A per _meta-033).

6. Optimum-coverage probe verdict

mt = "marian"
# vendor: feature-extraction, feature-extraction-with-past, text2text-generation, text2text-generation-with-past
# after_winml: identical (no override; pure-vendor coverage)
# added_by_winml: []

Verdict: VENDOR-COVERED on text2text-generation (composite expansion → encoder = feature-extraction, decoder = text2text-generation). Effort L0★ confirmed. Per winml config --task translation, the user-facing task translation correctly composite-expands to the two sub-tasks; the decoder recipe's task: text2text-generation is the canonical sub-task name per _meta-028.

7. Claimed (Effort, Goal, Outcome) tier

• Effort = L0★ (recipe-only; one winml config invocation per checkpoint, no hand-edits beyond _status removal which was never needed here)
• Goal = L2-encoder (L0 + L1-CPU PASS on both halves; L2 encoder cosine=1.0; L2 decoder DEFERRED-HARNESS per _meta-018)
• Outcome = L0 (recipe + finding append + this report; no source code; no feature-gap issues filed for this PR — the open feature gap "ship a winml.eval.compare_pt_onnx helper" is captured under marian-005 gotchas but is methodology-scope)

8. Goal-ladder verdict table (per _meta-018, per-half per _meta-020)

Short-circuit honored: no FAIL anywhere. L3 CLI-BLOCKED + L2-decoder DEFERRED-HARNESS do not halt the march per _meta-018. The honest ceiling is L2-encoder PASS.

Diligence ladder (_meta-037): not invoked — no BLOCKED-style verdict required ladder walk; the two BLOCKED verdicts (L1-non-CPU + L3) are host/CLI capability gaps documented in existing findings, not failed attempts.

9. Methodology-evolution declaration (per _meta-031)

No NEW methodology friction in this PR. The composite-recipe pattern + task=translation routing + decoder L2 harness gap were all captured during iter-5 (marian-003..005); they ship as separate _meta-NNN findings on the skills branch under _meta-019..030. Triggers:

• (1) CLI surprise — none.
• (2) Doc-code drift — none.
• (3) Silent-failure mode — none observed (cross-attention alias direct-name-match per _meta-025).
• (4) New verdict shape — DEFERRED-HARNESS was new during iter-5 but is now in the vocabulary.
• (5) Reviewer-found gap — pending reviewer pass.
• (6) Effort mis-estimate — none.
• (7) PR-mining discovery — none beyond _meta-019..030 already shipped.

Reviewer should confirm "no methodology friction observed" rather than REQUEST_CHANGES on absence per _meta-031 anti-trigger.

Reviewer hand-off package — Step 6 9-item self-check

1. Recipe files — §1 ✓
2. README row — §2 ✓ (to add in this PR)
3. Build output dir + artifact inventory — §3 ✓
4. Build log — §4 ✓
5. Appended findings — §5 ✓
6. Optimum-coverage probe verdict — §6 ✓
7. Claimed (Effort, Goal, Outcome) tier — §7 ✓
8. Goal-ladder verdict table — §8 ✓ (per-half, composite-expanded)
9. Methodology-evolution declaration — §9 ✓