Improve LoCoMo source-turn recall confidence

[okJiang]

Summary

Adds a guarded source-turn evidence bonus for LoCoMo recall confidence. The boost applies only to insight candidates with same-speaker source turns, exact/general query shapes, and bounded confidence thresholds; it skips duration, frequency, modal/inference, location-like, non-insight, and already-high-confidence candidates.

Benchmark Evidence

• Baseline run: 20260526T161100
• Baseline artifact: /Users/jiangxianjie/code/okjiang/harness/mem9-locomo/20260526T161100/benchmark-results.json
• Baseline Overall LLM (micro): 0.6396103896
• Candidate run: 20260526T213228
• Candidate artifact: /Users/jiangxianjie/code/okjiang/harness/mem9-locomo/20260526T213228/benchmark-results.json
• Candidate Overall LLM (micro): 0.6487012987
• Absolute delta: +0.0090909091
• Promotion target: >= 0.6446103896
• Counts/config: 1986 total QA, 1540 LLM-judged QA, no sample/category filters, model qwen3.6-plus, dataset /Users/jiangxianjie/code/mem9-locomo-workspace/mem9-benchmark/locomo/data/locomo10.json
• Ordered-row LLM flips versus baseline: 59 wins / 45 losses / 1436 ties

Category LLM deltas versus baseline:

• Cat1: 0.3971631206 (+0.0000000000)
• Cat2: 0.7445482866 (+0.0000000000)
• Cat3: 0.3958333333 (+0.0208333333)
• Cat4: 0.7253269917 (+0.0142687277)
• Cat5: N/A

Known non-primary regressions:

• Overall F1: 0.6123243694 (-0.0016045183)
• Overall Evidence Recall: 0.6888497265 (-0.0064467109)
• Cat5 Evidence Recall: 0.6580717489 (-0.0269058296)

Source Binding

• Candidate product branch: codex/locomo-20260526T213228-source-turn-boost-v2
• Candidate product commit: ca32790c0368c0e775757b23c55f8998e6c93b32
• Comparable baseline product HEAD: bb0a522194275c2c425f920d5928e5e579d9c612
• Baseline observed mem9 origin/main: fd0e7d38f19d3d84550978e43922284d391182e2
• Benchmark repo commit: 69d0d9c8fe665259bd3234fe9d16a256af266d34

Local Checks

• git diff --check
• go test ./internal/handler -run 'SourceTurn|RecallConfidence|RecallCandidateOptions|AnswerEvidence|QuestionSpecific'
• go test ./internal/handler ./internal/service

Integrity Notes

The benchmark dataset, gold answers, judge prompts, scoring logic, and category labels were not changed. The run used the Harness ./locomo.sh full benchmark with all samples and all categories.