feat(engine): community state as context, measured and on by default (JEV-30)

[ohernandezdev]

Experiment D (JEV-30): does a compact community-state line beside the conversation window change verdicts.

• Pre-registered criterion in benchmark/community_state/run.py, pushed in 9948223 before the first paid request.
• 178 constructed rows (spam, harassment, scam: paired incident/calm halves, bystanders during incidents, busy normal chat, controls), blind second labeller.
• Result (benchmark/community_state/REPORT.md): the whole line passes, pooled gain +32 points (97.5% lower end +20), no bystander or busy-chat verdict moved, busy-chat drift +0.036 under the +0.10 guard; +41 billed tokens (1.019x). The gain is recall, mostly spam copies; it does not clear calm lines; harassment not shown.
• Engine: Message.community (a CommunityState of counts) rendered by render_state into community_state, at most 50 estimated tokens, part of the cache key and the token estimate, behind JEVMOD_FULL_CONTEXT (default on). No caller fills it yet, so every request today is byte for byte unchanged. npm has no per-message context, no change.

$0.099 spent.

[github-actions]

Review summary

🔴 Someone should look at this before it is merged

feat(engine): community state as context, measured and on by default (JEV-30)
by @ohernandezdev · 14 files changed · 3099 lines changed

Why

• It is large: 3099 lines changed, and the limit for merging without a person is 400
• If this change is wrong the damage is medium: shared logic that several features rely on
• It may change more than its title says (56%)
• A mistake here might go unnoticed (51%)
• The worst case cannot be ruled out (16%)
• It may touch money

What the automatic review found

Question	Answer
If this change is wrong, how bad is it?	🟠 Medium -- shared logic that several features rely on
Does the title describe the whole change?	⚠️ Not confident either way (does what it says 93%, extra changes 56%)
Does it touch money, accounts or personal data?	⚠️ Not confident either way (61%)
Could a mistake here break things quietly?	⚠️ Not confident either way (51%)
Does it include tests?	✅ Yes

Each percentage is how likely the review thinks that answer is -- not a grade for the code.
Nothing was merged automatically -- this review only leaves a comment.

Raw scores

dimension	max across files
risk_level	1.99 / 3
worst_case_risk	0.16
diff_matches_title	0.93
hidden_scope	0.56
silent_failure	0.85
silent_failure_weighted	0.51
tests_expected	0.96
touches_money	0.61
touches_accounts	0.24
touches_personal_data	0.38
sensitive_area	0.61

114899 input tokens · cost of this run: less than a cent ($0.004826)

13 files reviewed.