Fixes a false-positive edge case (snipe#52) where sellers with 500+
lifetime feedback were hard-flagged as established_bad_actor when the
12-month ratio dipped below 80% — even though the 12-month window may
cover only a small recent sample relative to lifetime history.
Changes:
- established_bad_actor hard filter now only fires for accounts with
20–500 lifetime feedback (unchanged behavior for moderate accounts)
- Accounts >500 feedback with ratio 60–80%: new declining_ratio soft flag
(composite score penalised but not zeroed, no hard block)
- Accounts >500 feedback with ratio <60%: still established_bad_actor
(catastrophically bad even for high-volume sellers)
- Two new constants: HARD_FILTER_BAD_RATIO_MAX_COUNT=500,
HARD_FILTER_BAD_RATIO_HIGH_THRESHOLD=0.60
Note: buyer-feedback-only accounts (lifetime buyer history inflating
feedback_count for new sellers) requires profile-page scraping to detect
properly — tracked in snipe#52 as medium-term work.
Tests: 22 passed
Add 10 new tests covering the three previously untested flag paths:
- zero_feedback: flag fires + composite capped at 35 even with all-20 signals
- long_on_market: fires at >=5 sightings + >=14 days; NOT at <5 sightings or <14 days
- significant_price_drop: fires at >=20% below first-seen; NOT at <20% or no prior price
- established_retailer: duplicate_photo suppressed at feedback>=1000; fires below threshold
Also fix datetime.utcnow() deprecation in aggregator._days_since() and test helper
— replaced with timezone-aware datetime.now(timezone.utc) throughout.
- aggregator: also check listing.condition against damage keywords so listings
with eBay condition "for parts or not working" flag scratch_dent_mentioned
even when the title looks clean
- aggregator: add "parts/repair" (slash) + "parts or not working" to keyword set
- trust/__init__.py: pass listing.condition into aggregate()
- 3 new regression tests (synthetic fixtures, 17 total passing)
- SearchView: extract DEFAULT_FILTERS const + resetFilters(); add "Clear filters"
button that shows only when activeFilterCount > 0 with count badge
- .env.example: document LLM inference env vars (ANTHROPIC/OPENAI/OLLAMA/CF_ORCH_URL)
and cf-core wiring notes; closes#17
Large retailers like Newegg legitimately reuse manufacturer stock photos
across listings. Duplicate photo hash is not a scam signal for sellers
with 1000+ feedback — suppress the red flag for them.
- Parallel execution: search() and get_completed_sales() now run
concurrently via ThreadPoolExecutor — each gets its own Store/SQLite
connection for thread safety. First cold search time ~halved.
- Pagination: SearchFilters.pages (default 1) controls how many eBay
result pages are fetched. Both search and sold-comps support up to 3
parallel Playwright sessions per call (capped to avoid Xvfb overload).
UI: segmented 1/2/3/5 pages selector in filter sidebar with cost hint.
- True median: get_completed_sales() now averages the two middle values
for even-length price lists instead of always taking the lower bound.
- Fix suspicious_price false positive: aggregator now checks
signal_scores.get("price_vs_market") == 0 (pre-None-substitution)
so listings without market data are never flagged as suspicious.
- Fix title pollution: scraper strips eBay's hidden screen-reader span
("Opens in a new window or tab") from listing titles via regex.
Lazy-imports playwright/playwright_stealth inside _get() so pure
parsing functions are importable without the full browser stack.
- Tests: 48 pass on host (scraper tests now runnable without Docker),
new regression guards for all three bug fixes.
Scraper can't fetch seller profile age without following each listing's
seller link. Using 0 as sentinel caused every scraped seller to trigger
new_account and account_under_30_days red flags erroneously.
- Seller.account_age_days: int → Optional[int] (None = not yet fetched)
- Migration 003: recreate sellers table without NOT NULL constraint
- MetadataScorer: return None for unknown age → score_is_partial=True
- Aggregator: gate age flags on is not None
- Scraper: account_age_days=None instead of 0
