v0.12.17 - 🔍 City Pulse Review Pipeline: launch to hardened, weight-aware methodology (v0.12.0–v0.12.17)

🚀 New: the City Pulse review pipeline

An admin-triggered AI agent that audits City Pulse's own operation — cost, translation caching, severity weights, status/trend usefulness, prompt quality — and proposes concrete fixes as a draft PR.

• 🧮 Deterministic reducer (review/reduce.py) turns days of raw debug logs + radar.db into a compact, cheap-to-read digest — no LLM spend on re-parsing boilerplate
• 🤖 Gemini reviewer (review/reviewer.py) reasons over the digest only, producing a human-readable report plus machine-readable proposed edits
• 🔀 CI-mediated draft PR — proposals never auto-apply; a GitHub Actions workflow turns accepted findings into a reviewable PR
• 🏷️ Every pulse run stamped with pulse_config_version so output-quality changes can be attributed to a specific methodology snapshot and compared version-over-version

🛡️ Hardened against real production data

Validated with a gold-standard methodology — an independent human read of real digests vs. the actual Gemini report — catching real bugs before trusting the pipeline's own output:

• 🐳 Fixed missing review/+prompts/ in the web Docker image, and a missing GEMINI_API_KEY on the web container
• 🔧 Fixed the reducer's service filter polluting pulse_hours with extraction/daily records
• 📅 Fixed a cost-reconciliation month-boundary double-count
• 📊 Split paid_churn/cache_churn so a zero-cost flapping alert can't be mistaken for a cost driver
• 📈 Added a status_distribution histogram so the reviewer stops tallying status counts by eye
• ✍️ Corrected reviewer guidance on false "never reaches clear" findings
• 🏷️ Renamed a naming-collision field (prompt_samples → prompt_sample_texts) that caused a hallucinated setting

🎯 Reviewer can now actually judge and propose severity weights

The reviewer had no repo access and no visibility into today's actual weight values or thresholds — it could only guess. Fixed:

• ⚖️ weight_tables — the live SEVERITY_WEIGHTS_DWD/SERVICE_WEIGHTS_RMV/etc. values, read straight off pulse_categories.py
• 📐 Per-hour baseline (mean/p25/p75) alongside each category's score — the actual threshold a status was judged against
• 📝 top_drivers now carry title/source/body, not just id+weight, so a proposed change can cite what an alert actually said

🔧 Pulse-system fixes

• ✂️ Cap prompt alerts by severity rank instead of sending every alert (140 alerts / 68KB → bounded, ranked)
• 🌐 Difflib-based similarity threshold before paying for a retranslation on a cosmetic body edit
• 🕵️ Explicit root-cause-naming instruction so the pulse LLM connects correlated disruptions to their shared cause

🖥️ Admin dashboard cleanup

• 🎨 Colored pill for the review service in the Gemini log
• 🧹 Trimmed redundant status dots (RAM/Load — already on the health chart; StaticEvents — can't fail; Cron — duplicates "Last poll")
• 🏷️ Renamed "Extraction" health check to "Gemini"
• 🔴 "Last poll" text turns red once it exceeds the configured stale threshold
• 💾 Added current disk usage display
• 🗑️ Added 30-day retention to review_debug/ — previously the only debug directory with none

🐛 Fixes

• 💶 Fixed cost.monthly_budget_eur → cost.monthly_budget key mismatch (admin budget always silently fell back to a hardcoded €10 default)
• 🗂️ Moved translation_min_change_ratio under translator: in config.yaml (it's a caching knob, not pulse methodology)