Keep a mixed document's parse metrics when the provider emits no layout

[boyang-zhang1]

Stacked on #133 (its commit is the first commit here, authorship preserved). Groundwork for #154. A document that carries both layout rules and parse rules routes to _evaluate_multi_task, and that path has three losses that only become reachable once such a document exists — which is exactly what #133 enables. This lands them first so #133 can go in without moving anyone's score in a way nobody expects.

No shipped dataset carries a mixed document today (data/layout.jsonl is 100% type: "layout", and the group key is (category, pdf)), so this changes no current leaderboard number. Verified by loading data/ and data/test/ before and after: identical.

1. A markdown-only provider scored 0 on a document it partly passed

create_layout_adapter_for_result sat outside the try block. For any provider with no layout output (openai, anthropic, gemini, reducto…) it raises, _evaluate_single turns the document into success=False, metrics=[], and _aggregate_metrics drops every metric of a failed result — including the parse rules that had just passed — before _is_infra_failure zero-pads it.

Reproduced on main with one layout rule + one present rule and a markdown-only parse result:

main: rules log "1/1 passed", then success=False n_metrics=0
      error='Evaluation error: Inference output is not LayoutOutput and no provider adapter matched.'
this: success=True  metrics=[('rule_pass_rate', 1.0), ('rule_present_pass_rate', 1.0)]

Only the default adapter's "not LayoutOutput" ValueError is caught — a real adapter bug still propagates and still fails the result, which is what _is_infra_failure exists to sort out (covered by a test).

When no adapter matches, the provider emitted nothing the layout task can score, so the layout half scores a genuine 0 — the same verdict _is_infra_failure reaches for a pure-layout document. An empty LayoutOutput tagged with a new LayoutDetectionModel.NONE sentinel is run through the evaluator, which emits the full metric set with its own denominators, so the document stays in the AP/F1/layout_rule_pass_rate denominators. Leaving it unscored (the first revision of this PR) let one perfect document plus one markdown-only document aggregate to avg_AP50 = 1.0.

When an adapter matched but found no elements, that is scoreable and is now scored instead of failed. The evaluator runs against the empty prediction set and returns a genuine 0 across the whole metric set, with the denominators it builds itself:

rule_pass_rate 1.0 {passed: 1, total: 1}     <- parse half intact
layout_rule_pass_rate 0.0 {passed: 0, total: 2}   <- localization + classification
AP50 0.0   mean_f1 0.0   num_ground_truth 1.0   unmatched_gt_elements 1.0

That branch previously appended "Could not extract layout from PARSE output" and failed the whole result.

2. The split parse test case lost the document's parse configuration

It was rebuilt with expected_markdown=None and default table settings, so a mixed document was scored without the markdown GT that drives text similarity / TEDS / GriTS, and with the title-strip and TRM fallback reset — a document flagged trm_unsupported: true was scored with the metric it was flagged unreliable for. It now inherits expected_markdown, allow_splitting_ambiguous_merged_tables, trm_unsupported, max_top_title_rows and metadata.

Separately, metadata was read only from LayoutDetectionTestCase, so source_dataset was silently None whenever the mixed rules arrived on a ParseTestCase — which is the case the sidecar loader already produces. It is read by attribute now, and a LayoutDetectionTestCase's own source_dataset field is preferred over the metadata copy.

3. rule_pass_rate was emitted twice into one result

The layout evaluator emits rule_pass_rate as an alias of the layout_rule_pass_rate right beside it, and RuleBasedMetric emits the same name. A mixed result carried both, so one document contributed two samples to avg_*, pooled unrelated denominators into micro_*, and the CSV export and detailed report — which flatten with {m.metric_name: m.value} — kept whichever came last, making the per-example view disagree with the rule detail pane under it.

The layout alias is dropped when the parse half already emitted that name. Not folded arithmetically: the parse entry carries the rule_results payload the report renders, and its value is a graduated score rather than passed / total, so combining them by counts would erase partial credit and drop the rule details. layout_rule_pass_rate keeps the layout half's number and subcounts.

Tests

tests/parse_bench/evaluation/test_multi_task_evaluation.py, 10 tests: the markdown-only case for both test-case classes, the two-document aggregate that must not read 100% when one document has no layout, the empty-layout-output case scoring a real zero with the evaluator's own denominators, a real adapter bug still failing the result, config and source_dataset inheritance from either carrier, and the alias-drop rules. ruff check, ruff format --check and the evaluation + test_cases + analysis suites (2596 tests) are clean.