כל משפט בעמוד הזה אפשר לפתוח ולבדוק.

PREV/report.json      analysis.overall = 75
POST-FIX/report.json  analysis.overall = 86

  speed    59  → 87
  mobile   87  → 100
  google   72  → 72
  trust    87  → 87

  page bytes  377006 → 10155   (377KB → 10KB)
  measured    2026-07-11T02:46:38.075Z → 2026-07-11T18:07:55.207Z

  still open (3):
    asset-cache    speed    med
    empty-shell    google   high
    soft-404       trust    med

שתי הבדיקות רצו באותו מנוע על השרת החי. מה שעוד לא תוקן — רשום.

ads.json  generated_at 2026-09-11T08:05:13Z

  keywords        121
  negatives       283
  ad_groups       4
  ads             13
  alerts          4

run-ads-ops.sh — the only Ads tools this job may call:
    GOOGLEADS_SEARCH_STREAM_GAQL
    GOOGLEADS_MUTATE_AD_GROUP_CRITERIA
    GOOGLEADS_MUTATE_AD_GROUP_ADS
  preflight fails the run if any of these is ever allow-listed:
    GOOGLEADS_MUTATE_CAMPAIGNS
    GOOGLEADS_MUTATE_CAMPAIGN_BUDGETS
    GOOGLEADS_MUTATE_BIDDING_STRATEGIES

**The trap that was avoided.** The obvious API route
(`MUTATE_CAMPAIGNS` -> `manual_cpc`) would have
been a disaster: every ad group carries `cpc_bid_micros:
10000` (**₪0.01**) — the placeholder Google
keeps while smart bidding runs — against an actual
clearing CPC of **₪5.63–9.07**. Manual CPC would
have bid one agora into a ₪7 auction and stopped delivery
entirely while the campaign still read
ENABLED. Always read ad-group `cpc_bid_micros` against
`average_cpc` before switching strategy.

המספרים כאן הם תחזוקה, לא תוצאה. הם מוכיחים שהמערכת רצה על חשבון אמיתי — לא שהיא הרוויחה.

$ grep -cE '^\s*assert\(' src/self-test.js   # calls only, not the helper's definition
  48

$ ls scenarios/transcripts/*.md | grep -v SUMMARY | wc -l
  14

  after-hours-inquiry.md  ambiguous-rambling.md  angry-complaint.md  booking-happy-avi-vardi.md  booking-happy-barbershop.md  competitor-question.md  emoji-only.md
  mixed-language.md  off-topic-politics.md  optout-mid-booking.md  price-ask-barbershop.md  price-ask-no-quote-avi.md  shabbat-refusal.md  tier2-offline-fallback.md

הבדיקות רצות אצלי במחשב לפני כל שינוי, ולא דרך שרת ציבורי.

site/data/status/book-0.json   (kesef-20s)

  judge.sources   248
  judge.pages     193
  judge.score     94 / 100
  first judge run 84 / 100  FAIL vs a 90 bar

  volumes 4, with audio 3:
    kesef-20s  score 94   audio yes
    dira       score 91   audio yes
    modaa      score 96   audio no
    rega       score 86   audio yes

הציון זז לשני הכיוונים: כרך מאוחר יותר ירד בציון אחרי סבב תיקון, וגם זה רשום.