Add global temperature analysis tasks

[ScuttleBot]

Adds 3 new data-analysis tasks using the global_temperature.csv dataset:

1. task_csv_temp_anomalies - Detect temperature anomalies (extreme months, z-score outliers, warmest/coldest years, year-over-year changes)
2. task_csv_temp_trend - Analyze temperature trends (linear regression, pre/post-1950 acceleration, milestone crossings)
3. task_csv_temp_decades - Compare decades (decade averages, transitions, variability, GISTEMP vs gcag source comparison)

All tasks use hybrid grading (60% automated, 40% LLM judge) with 180s timeout.

Closes #219, Closes #221, Closes #222

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Solid set of data-analysis task definitions. The grading scripts use regex-based pattern matching appropriately, expected values are well-documented, and the hybrid grading weights (60% automated / 40% LLM judge) make sense for this type of open-ended analysis task.

Files Reviewed (4 files)

• tasks/manifest.yaml
• tasks/task_csv_temp_anomalies.md
• tasks/task_csv_temp_trend.md
• tasks/task_csv_temp_decades.md

---

Fix these issues in Kilo Cloud

---

_{Reviewed by claude-4.6-sonnet-20260217 · 147,738 tokens}

[ScuttleBot]

🧪 PR Test Started

Instance: 155.138.235.245 (Vultr vc2-2c-4gb, ATL)
Branch: tasks/csv-temperature
Tasks:

• task_csv_temp_anomalies
• task_csv_temp_trend
• task_csv_temp_decades

Models (running in parallel):

#	Model
1	openrouter/anthropic/claude-opus-4.6
2	openrouter/openai/gpt-5.4
3	openrouter/google/gemini-3.1-pro-preview

ETA: ~15-20 minutes (3 models × 3 tasks, 180s timeout each)

Automated PR test by ScuttleBot 🦀

[ScuttleBot]

🦀 PR Test Results — Global Temperature Tasks

Instance: 155.138.235.245 (Vultr vc2-2c-4gb, ATL)
Branch: tasks/csv-temperature
Grading: Hybrid (60% automated + 40% LLM judge)
Duration: ~28 min total (sequential, agent concurrency issue forced non-parallel)

Overall Scores

Model	Anomalies	Trend	Decades	Overall
openrouter/google/gemini-3.1-pro-preview	90.7%	96.0%	100.0%	95.6%
openrouter/anthropic/claude-opus-4.6	99.6%	97.6%	95.6%	97.6%
openrouter/openai/gpt-5.4	99.0%	98.0%	95.6%	97.5%

Task Breakdown

task_csv_temp_anomalies

Model	Score	Time	Tokens	Notes
Gemini 3.1 Pro	90.7%	64s	91,747	yoy_changes partial (0.5)
Claude Opus 4.6	99.6%	81s	83,089	Near-perfect
GPT-5.4	99.0%	68s	92,625	Near-perfect

task_csv_temp_trend

Model	Score	Time	Tokens	Notes
Gemini 3.1 Pro	96.0%	71s	79,430	All automated checks passed
Claude Opus 4.6	97.6%	90s	101,337	All automated checks passed
GPT-5.4	98.0%	53s	55,705	All automated checks passed, fastest

task_csv_temp_decades

Model	Score	Time	Tokens	Notes
Gemini 3.1 Pro	100.0%	79s	112,383	Perfect score
Claude Opus 4.6	95.6%	124s	108,781	Slowest on this task
GPT-5.4	95.6%	58s	59,981	Most efficient (fewest tokens)

Observations

1. All 3 tasks work well — every model scored 90%+ on all tasks, confirming the tasks are solvable and the grading is reasonable
2. Task quality looks solid — the hybrid grading (automated + LLM judge) produces nuanced scores that differentiate between partial and full solutions
3. Gemini partial on yoy_changes in anomalies task — the year-over-year changes check scored 0.5 (partial credit), suggesting the automated grader may need a slightly broader acceptance range, or Gemini's formatting differed from expected
4. Token efficiency varies widely — GPT-5.4 used 2x fewer tokens than Gemini/Claude on trend and decades tasks while scoring comparably
5. All models completed well within the 180s timeout — longest execution was Claude at 124s on decades

Verdict

✅ Tasks are ready to merge. All three produce meaningful, differentiating results across models with no failures or edge cases.

Tested by ScuttleBot 🦀