recipe(pix2text-mfr): pin VED template as KNOWN-BROKEN regression coverage

[ssss141414]

PR: breezedeus/pix2text-mfr — VED image-to-text recipe pair (KNOWN-BROKEN regression-pin)

Iter: 6 (validated-negative shipped iter-5 as vision-encoder-decoder-003; this PR pins the recipes with _status markers per _meta-014 convention)
Producer: main agent (2026-06-23)
Claimed tier: (Effort = L0★, Goal = L0-FAIL-UPSTREAM, Outcome = L0)

Summary

This PR ships the breezedeus/pix2text-mfr image-to-text recipe pair as a KNOWN-BROKEN regression-pin per _meta-014 _status marker convention. The recipes are structurally correct VED templates (would build any standard HF VisionEncoderDecoderModel checkpoint with covered inner decoder); they are blocked at the upstream HF repo-layout layer, not at the winml export layer. Each recipe carries a top-level _status: "BROKEN — DO NOT USE. ..." field documenting the exact failure and the recipe's regression-coverage purpose.

This PR is shipped as a contribution because:

1. The recipes record the canonical L0★ VED template for any future contributor who needs a starting point.
2. The _status marker pattern is itself the regression-test for WinMLBuildConfig.from_dict correctly ignoring unknown top-level keys (positive control verified in iter-5 on opus-mt-fr-en encoder build with _status field present — build SUCCEEDED).
3. The failure class is upstream (HF repo only ships pre-exported .onnx files, no PyTorch weights for AutoModel.from_pretrained), so dropping the recipe entirely loses the diagnostic value.

No source-code changes.

1. Recipe files

• examples/recipes/breezedeus_pix2text-mfr/image-to-text_fp16_encoder_config.json
• examples/recipes/breezedeus_pix2text-mfr/image-to-text_fp16_decoder_config.json

Both recipes carry "_status": "BROKEN — DO NOT USE. winml build fails at fetch with \breezedeus/pix2text-mfr does not appear to have a file named pytorch_model.bin, model.safetensors, tf_model.h5, model.ckpt or flax_model.msgpack.`. The HF repo stores weights in a non-standard layout. Recipe is structurally correct and would build any standard VED checkpoint; pinned here as regression coverage. See research/adding-model-support/model_knowledge/vision_encoder_decoder.json finding vision-encoder-decoder-003."`

2. README index row

examples/recipes/README.md — row to add for breezedeus/pix2text-mfr | image-to-text | composite (encoder + decoder) | recipe pair | **BROKEN (upstream HF repo layout)**.

3. Build output directory + artifact inventory

temp/ved_build/ (gitignored) — EMPTY for this checkpoint. The encoder build attempt aborts at the fetch stage:

Error: Build failed: breezedeus/pix2text-mfr does not appear to have a file
named pytorch_model.bin, model.safetensors, tf_model.h5, model.ckpt or
flax_model.msgpack.

No model.onnx, no model.onnx.data, no analyze_result.json, etc. — the failure precedes the export pipeline.

External-data layout check (_meta-023): N/A (no artifacts produced).

4. Build log

The recipe _status field IS the build-log substitute — see §1. Stderr captured in vision-encoder-decoder-003 mechanism_notes.

5. Appended findings

Per-model — model_knowledge/vision_encoder_decoder.json

• vision-encoder-decoder-001 — pre-build coverage (generic VED export covers any standard HF VED checkpoint).
• vision-encoder-decoder-002 — REFINEMENT: Optimum natively covers VED image-to-text; winml ADDS encoder/decoder split overrides.
• vision-encoder-decoder-003 — VALIDATED-NEGATIVE for breezedeus/pix2text-mfr (this PR's primary basis).
• vision-encoder-decoder-004 — VED template confirmed reusable for nlpconnect/vit-gpt2-image-captioning (the recommended canonical L0★ VED reference instead of the broken one here).

Skill-meta

No new _meta-NNN findings in this PR.

6. Optimum-coverage probe verdict

mt = "vision-encoder-decoder"
# vendor: feature-extraction, image-to-text, image-to-text-with-past
# after_winml: + image-feature-extraction encoder override + text2text-generation decoder override
# added_by_winml: ["image-feature-extraction", "text2text-generation"]  (composite split)

Verdict: VENDOR-COVERED on image-to-text (composite splits into image-feature-extraction encoder + text2text-generation decoder via winml overrides per vision-encoder-decoder-002). Effort L0★ is correct CLASSIFICATION; in PRACTICE this specific checkpoint is upstream-blocked.

7. Claimed (Effort, Goal, Outcome) tier

• Effort = L0★ (recipe-only template; structurally identical to vit-gpt2-image-captioning)
• Goal = L0-FAIL-UPSTREAM (build fails at fetch — not a winml gate; the recipe + _status field together encode "we tried, here's why it can't proceed")
• Outcome = L0 (recipe + finding append + this report)

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

Half	Tier	Verdict	Evidence
encoder	L0	FAIL-UPSTREAM	winml build aborts at fetch: HF repo breezedeus/pix2text-mfr ships only encoder_model.onnx + decoder_model.onnx (pre-exported ONNX) + config.json / tokenizer* / preprocessor_config.json. NO pytorch_model.bin / model.safetensors / tf_model.h5 / model.ckpt / flax_model.msgpack. AutoModel.from_pretrained requires one of these to materialize the PyTorch graph that winml build then traces.
decoder	L0	FAIL-UPSTREAM	Same root cause — the fetch failure happens before encoder/decoder split.
L1..L3 (both halves)	NOT-REACHED	Per _meta-018, FAIL halts the march. Lower tiers are unreachable until L0 is unblocked.

Short-circuit honored: L0 FAIL-UPSTREAM is the FIRST FAIL verdict in the ladder, halts the march. The _status marker on the recipe is the artifact-of-record for this halted state.

Diligence ladder (_meta-037) — invoked and recorded:

1. Re-read vision_encoder_decoder.json — vision-encoder-decoder-003 fully documents the failure mode; no workaround documented because the gate is upstream.
2. PR-mine — no PR has unblocked this checkpoint family; vit-gpt2 sibling (vision-encoder-decoder-004) works on standard-layout checkpoints.
3. Re-run winml config — already succeeded; winml build is where it fails. winml config produced the recipe pair correctly from the HF config alone (doesn't need weights).
4. --ep-options — N/A (failure is at fetch, not at EP).
5. value_range / shape pinning — N/A (failure is at fetch, not at export).
6. Custom Python harness — N/A (we don't have a PyTorch model to compare against; the upstream gate is precisely "no PyTorch model in repo").
7. Re-verify upstream repo layout (2026-06-23): HfApi.list_repo_tree('breezedeus/pix2text-mfr', recursive=True) returns: .gitattributes, README.md, config.json, decoder_model.onnx, encoder_model.onnx, generation_config.json, preprocessor_config.json, special_tokens_map.json, tokenizer.json, tokenizer_config.json. NO PyTorch weights, only pre-exported ONNX. The failure mode is unchanged from iter-5 (vision-encoder-decoder-003).

Feature gap from step 7: winml build could conceivably support a "fetch pre-exported ONNX from HF instead of PT→ONNX export" path for repos like this. Captured under vision-encoder-decoder-003 feature_gaps_filed[]. Until then, this checkpoint stays BROKEN.

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

No NEW methodology friction in this PR. The _status marker convention (_meta-014) was the iter-5 finding that enabled shipping known-broken recipes; this PR uses it as designed. Triggers:

• (1) CLI surprise — none.
• (2) Doc-code drift — none.
• (3) Silent-failure mode — none (failure is loud and reproducible).
• (4) New verdict shape — FAIL-UPSTREAM is already covered by the FAIL slot in _meta-018 vocabulary (the -UPSTREAM suffix is descriptive, not a new verdict).
• (5) Reviewer-found gap — pending.
• (6) Effort mis-estimate — none.
• (7) PR-mining discovery — none.

Reviewer should confirm "no methodology friction observed" per _meta-031 anti-trigger. The diligence ladder application IS the methodology working as designed — recipe ships with a documented FAIL-UPSTREAM verdict + feature_gap entry, not a silent failure.

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

1. Recipe files — §1 ✓ (both carry _status marker)
2. README row — §2 ✓ (to add in this PR; row marks the recipe as BROKEN)
3. Build output dir + artifact inventory — §3 ✓ (empty by design, reason documented)
4. Build log — §4 ✓ (substituted by _status field — exact stderr captured in vision-encoder-decoder-003 mechanism_notes)
5. Appended findings — §5 ✓
6. Optimum-coverage probe verdict — §6 ✓
7. Claimed tier — §7 ✓ (Goal = L0-FAIL-UPSTREAM honestly)
8. Goal-ladder verdict table — §8 ✓ (FAIL-UPSTREAM halts march per _meta-018)
9. Methodology-evolution declaration — §9 ✓