Add OpenDataLoader parity coverage and Java-core auto rescue

[JamesbbBriz]

Summary

• Add OpenDataLoader pipeline parity contracts: processor matrix, stage order, heuristic ownership, behavior-family buckets, and full200 gate metadata.
• Gate Java-core preset=auto prediction so readable Java/PDFBox output stays canonical, while sparse output can route through Rust/MNN OCR rescue.
• Keep model workers warm across OpenDataLoader prediction batches and prepare local PP-OCRv5 MNN cache when needed.
• Generate concrete low-score-buckets.json artifacts from evaluator output so full200 gaps are machine-readable by metric bucket and matrix-aligned behavior family.

Full200 Verification

Latest release full200, Java-core auto + Rust/MNN rescue:

artifact: third_party/opendataloader-bench/prediction/doctruth-java-core-20260627T190836Z/full200
parsed:   200/200
failed:   0
overall:  0.781875
nid:      0.900985
teds:     0.736174
mhs:      0.492119
latency:  108.615072 ms/doc
rss:      start 7 MB, peak 71 MB, end 71 MB
python/torch/docling residency: false
model routes: 1 OCR route, 01030000000141
low-score artifact: low-score-buckets.json
metric buckets: heading 58, reading-order 23, table 7, overall 22, missing 0
behavior buckets: text-noise 22, two-column 23, sidebar 23, heading 58, bordered-table 7, borderless-table 7, ocr-rescue 0
classification basis: metric_proxy until richer layout tags are available

Earlier release full200, deterministic Java-core lite baseline:

artifact: third_party/opendataloader-bench/prediction/doctruth-java-core-20260627T174900Z/full200
parsed:   200/200
failed:   0
overall:  0.779731
nid:      0.898148
teds:     0.736174
mhs:      0.489455
latency:  64.569802 ms/doc
rss peak: 19 MB
python/torch/docling residency: false
model routes: 0

Test Plan

cargo fmt --manifest-path runtime/doctruth-runtime/Cargo.toml -- --check
cargo test --manifest-path runtime/doctruth-runtime/Cargo.toml --test benchmark_corpus_contract opendataloader_ -- --nocapture
cargo test --manifest-path runtime/doctruth-runtime/Cargo.toml --test opendataloader_parity_matrix_contract -- --nocapture
cargo test --manifest-path runtime/doctruth-runtime/Cargo.toml
sh scripts/smoke-doctruth-runtime.sh
sh scripts/smoke-doctruth-runtime-model-worker.sh
sh scripts/smoke-doctruth-runtime-benchmark-corpus.sh
git diff --check
DOCTRUTH_RUNTIME_BUILD_PROFILE=release DOCTRUTH_OPENDATALOADER_PRESET=auto sh scripts/run-opendataloader-java-core-parity.sh --full200

CI Note

The GitHub CI build (25) job is still red on mvn -B -ntp spotless:check checkstyle:check. I reproduced the same 117-file Spotless failure on main in a temporary worktree, so it is a baseline formatting issue rather than a regression from this PR. I am intentionally not mixing a repo-wide Java formatting sweep into this parser parity PR.

Notes

• OpenDataLoader remains a behavior reference and benchmark surface, not a canonical output schema or production fallback.
• TrustDocument remains canonical.
• low-score-buckets.json now separates raw metric buckets from matrix-aligned behavior buckets. Behavior buckets are metric-proxy classifications until the evaluator consumes richer layout tags.

[JamesbbBriz]

CI note: the current failing build (25) job is from mvn -B -ntp spotless:check checkstyle:check reporting 117 Java formatting violations. I reproduced the same failure on main at 9df5f96, with the same 117-file Spotless baseline issue. I am not mixing a repository-wide Java formatting sweep into this parser parity PR. The parser/Rust validation and full200 gates for this branch have passed locally.

[JamesbbBriz]

Updated after closing the concrete low-score artifact gap.

Latest head: 04231a4 feat: write opendataloader low-score bucket artifacts

Fresh release full200, Java-core auto + Rust/MNN rescue:

artifact: third_party/opendataloader-bench/prediction/doctruth-java-core-20260627T185345Z/full200
parsed:   200/200
failed:   0
overall:  0.781875
nid:      0.900985
teds:     0.736174
mhs:      0.492119
latency:  128.819003 ms/doc
rss:      start 7 MB, peak 60 MB, end 60 MB
python/torch/docling residency: false
model routes: 1 OCR route, 01030000000141
low-score artifact: low-score-buckets.json
low-score cases: 76/200; heading 58, reading-order 23, table 7, overall 22, missing 0

Fresh verification run locally:

cargo fmt --manifest-path runtime/doctruth-runtime/Cargo.toml -- --check
cargo test --manifest-path runtime/doctruth-runtime/Cargo.toml --test benchmark_corpus_contract opendataloader_ -- --nocapture
cargo test --manifest-path runtime/doctruth-runtime/Cargo.toml
sh scripts/smoke-doctruth-runtime.sh
sh scripts/smoke-doctruth-runtime-model-worker.sh
sh scripts/smoke-doctruth-runtime-benchmark-corpus.sh
git diff --check
DOCTRUTH_RUNTIME_BUILD_PROFILE=release DOCTRUTH_OPENDATALOADER_PRESET=auto sh scripts/run-opendataloader-java-core-parity.sh --full200

CI note remains: build (25) is red on the repo-wide Java Spotless/checkstyle baseline. I reproduced the same 117-file Spotless failure on main; this PR does not add that failure, and I did not mix a full Java formatting sweep into this parser parity branch.