Dataset expansion: +2.2k targeted markdown rows, 9 new detection labels

[bfirestone]

Summary

This PR closes a critical dataset gap in the training corpus and adds 9 new detection labels for popular embedded languages. Follow-up to the detection label audit work.

The Gap

The prior corpus had zero markdown rows containing rust code blocks and only 9 containing go. The detection head literally cannot learn to detect rust-in-markdown when no such examples exist. Similar gaps existed for swift (6), kotlin (2), lua (2), and objc (3).

The Fix

Adds --content-filter support to pull_real_data to target specific languages when streaming The Stack markdown. Uses this to pull ~2,200 markdown documents containing embedded code blocks for 17 languages.

Coverage improvements (markdown rows containing each language):

Language	Before	After	Δ
rust	0	340	∞
go	9	295	+286
swift	6	173	+167 (new label)
kotlin	2	150	+148 (new label)
lua	2	142	+140 (new label)
r	1	104	+103 (new label)
objc	3	95	+92
graphql	0	82	+82 (new label)
ruby	19	82	+63 (new label)
csharp	42	64	+22 (new label)
powershell	29	37	+8 (new label)
php	17	25	+8 (new label)

New Detection Labels (9)

Added to DETECTION_LABELS with explicit definitions in SYSTEM_PROMPT:

csharp, powershell, ruby, php, swift, kotlin, r, lua, graphql

Each definition includes positive signals (distinctive syntax/keywords) and anti-signals (common confusions with existing labels). The r, swift, and go definitions have a Name alone in prose is not sufficient caveat to prevent prose-mention false positives.

Validation

Ran a 3-model 5k stratified audit after the expansion using gemini-3-flash, sonnet-4.6, and gpt-5.4-mini. Results:

• rust cross-fires: 3 → 29 (+867%)
• go cross-fires: 5 → 20 (+300%)
• All 9 new labels fire with 100% recall on ground-truth markdown rows
• Inter-annotator agreement ≥0.999 across 3 models for all new labels
• Zero ruby/rust false positives from the r regex (critical test)

Commits

1. d33416e — Targeted pull infrastructure: --sub-type, --target, --content-filter, --source-label, --max-skips-multiplier flags + click wrapper + 49 tests including R-regex precision tests
2. 4cffd46 — 9 new detection labels with SYSTEM_PROMPT definitions and caveats
3. 7c34e00 — Dataset expansion (2,186 new rows) + audit_detection_labels.py analysis script + DVC hash update

Data

• golden_train.parquet DVC hash: 3cd90fbd… → c15c01ff…
• Corpus size: 88,440 → 90,626 rows (+2,186 net new)
• Pushed to DVC remote (B2)
• Old corpus preserved as golden_train.r4.parquet for rollback

Known Issue (Filed Separately)

Discovered during this work: the existing golden_train.parquet is not a fixed point under the current dedup pipeline — re-running dedup on it alone removes 8,053 rows. Worked around by skipping cross-corpus dedup and using only exact-text deduplication for the expansion. Filed as arc bug textclf-1dd1.013xtr for separate investigation.

Test plan

• 49 new unit tests for pull infrastructure including R-regex precision (all pass)
• Dry-run annotation with updated 40-label prompt (verified 40 det columns)
• 3-model 5k stratified audit on expanded corpus (verified gap closure)
• Ground-truth recall check (100% on markdown rows containing each language)
• Feature sanity check on new rows (no NaN/Inf/NULL)
• Content spot check on 10 random new rows (all legitimate documentation)
• Internal near-duplicate cleanup (12 pairs removed)
• DVC push verified (6 files to B2)