feat: Paper results pipeline and robustness improvements

[Mattdl]

Summary

Adds end-to-end support for reproducing paper benchmark results — from running experiments to generating publication-ready LaTeX tables — while fixing several robustness issues discovered during large-scale multilingual evaluation runs.

Changes

New: Paper results pipeline

• examples/run_paper_results.py — Script to run the full benchmark suite across multilingual and monolingual model configurations (BM25, JobBERT-v2/v3, Qwen3, ConTeXTMatch, CurriculumMatch).
• examples/generate_paper_table.py — Script to load saved results and generate a formatted LaTeX comparison table with model grouping, short names, bold-best highlighting, and optional dataset count rows.
• src/workrb/metrics/reporting.py — New format_results_latex() function (~290 lines) that builds a complete LaTeX table environment from multiple BenchmarkResults, with support for model groups, \midrule separators, column renaming/reordering, \resizebox, and per-group dataset count rows.
• src/workrb/results.py — New BenchmarkResults.get_dataset_counts() method to report how many datasets contribute to each task group or task score after language filtering.

New: Duplicate handling in ranking datasets

• DuplicateStrategy enum (ALLOW, RAISE, RESOLVE) replaces the old boolean allow_duplicate_queries / allow_duplicate_targets flags in RankingDataset.
• RESOLVE mode (new default) deterministically deduplicates:
  • Targets: keeps first occurrence and remaps all indices.
  • Queries: merges target_indices via set union for identical query texts.
• All ranking tasks (melo, mels, skillnorm, etc.) updated to use the new API.
• tests/test_duplicate_strategy.py — 210-line test suite covering RAISE, ALLOW, and RESOLVE strategies for both queries and targets.

New: Graceful handling of unsupported dataset configs

• DatasetConfigNotSupported exception — New exception type in base.py for datasets that dynamically produce 0 queries or targets (e.g., an ESCO language/version lacking skill alternatives).
• Task._load_datasets() catches DatasetConfigNotSupported and logs a warning instead of crashing, then updates self.dataset_ids to reflect only successfully loaded datasets.
• RankingDataset validation raises DatasetConfigNotSupported when query or target count is 0.

New: ConTeXTMatch query batching for large ranking tasks

• ConTeXTMatchModel._compute_rankings() now scores queries in configurable chunks (scoring_batch_size, default 32) to prevent OOM from the (num_queries, num_targets, seq_len) intermediate tensor.
• Targets are encoded once and reused across all query chunks.
• Renamed batch_size parameter to encode_batch_size for clarity.
• tests/test_models/test_contextmatch_model.py — 108-line test suite for the ConTeXTMatch model including batching behavior.

Fix: Version-dependent ESCO language support

• ESCO.get_supported_languages(version) returns the correct language set per major.minor version (e.g., Icelandic/Norwegian/Arabic/Ukrainian only available from v1.1+).
• All ranking tasks now call ESCO.get_supported_languages(self.esco_version) instead of the static SUPPORTED_ESCO_LANGUAGES tuple.

Fix: Miscellaneous

• Cast prediction matrices to .float() before .numpy() to avoid dtype errors with bfloat16 models.
• Use sorted(set(...)) instead of list(set(...)) for deterministic index ordering.

Files changed (22)

Area	Files
Examples	examples/run_paper_results.py, examples/generate_paper_table.py
Metrics & Reporting	src/workrb/metrics/reporting.py, src/workrb/metrics/__init__.py, src/workrb/metrics/classification.py, src/workrb/metrics/ranking.py
Models	src/workrb/models/bi_encoder.py
Results	src/workrb/results.py
Tasks (core)	src/workrb/tasks/abstract/base.py, src/workrb/tasks/abstract/ranking_base.py, src/workrb/tasks/__init__.py
Tasks (ranking)	job2skill.py, skill2job.py, skill_extraction.py, jobnorm.py, skillnorm.py, melo.py, mels.py
Data	src/workrb/data/esco.py
Tasks (classification)	src/workrb/tasks/classification/job2skill.py
Tests	tests/test_duplicate_strategy.py, tests/test_models/test_contextmatch_model.py

Test plan

• Run tests/test_duplicate_strategy.py — validates RAISE, ALLOW, and RESOLVE strategies for both query and target deduplication
• Run tests/test_models/test_contextmatch_model.py — validates ConTeXTMatch encoding and batched scoring
• Run existing test suite to confirm no regressions in ranking/classification tasks
• Verify examples/run_paper_results.py executes end-to-end with at least one model
• Verify examples/generate_paper_table.py generates valid LaTeX output from saved results

Breaking changes

• RankingDataset.__init__ signature changed: allow_duplicate_queries / allow_duplicate_targets replaced by duplicate_query_strategy / duplicate_target_strategy (enum-based). Any external callers using the old boolean flags will need to update.
• ConTeXTMatchModel.encode() parameter renamed from batch_size to encode_batch_size.