v0.1.0 — search results and vehicle detail pages

First release. Two modes over autotrader.com's own page store, three browser
engines over one shared page loop, and the 2Captcha Scraper API client for
both modes.

Before your first run: this site serves its pages only to a US
residential exit with a headful browser (measured 2026-09-24; the
README has the table). The engines are headful by default and want
xvfb-run on a server, and a run needs --proxy or AUTOTRADER_PROXY.

Added

• --mode listing: one search's organic results, one row per car (50
  columns): price with fees, pre-fee price and fees, MSRP, KBB fair price
  and range, deal rating, VIN, year/make/model/trim, condition, mileage,
  body, colours, fuel, drivetrain, engine, transmission, mpg, days on site,
  stock number, photo, and the dealer or private seller with rating,
  address and distance.
• --mode vehicle: one detail page per listing id (56 columns): everything
  above plus features, the seller's description, open recalls, KBB owner
  reviews and the photo count. Ids from --vehicle-id, a detail --url, or
  --from-listing with a listing run's JSON.
• A search from --url (with the site's own filters kept) or from
  --make/--model/--zip/--radius/--listing-type; --sort over the seven
  orderings the site was measured to honour; --page-size 25|100.
• The sidecar records the site's own count, the pages it serves, the
  300-result cap (capped_by_site, reachable_max), the search the site
  says it ran (site_query), and in vehicle mode the listings that have
  gone (vehicles_gone).

Measured, and built in

• Headful and US residential, both. Headless (old and new) refused from
  a US residential exit; headful refused from a datacentre. Headful is the
  default, and a missing display is said before any launch.
• A slug the site does not know widens the search silently
  (f-150 → every used Ford, HTTP 200). The served path is compared with
  the requested one; a widened search writes nothing, exit 2.
• The site serves 300 results per search, and a page past the end is
  its LAST page re-served. The run ends on the page number the site says it
  rendered.
• Paid placements are not rows; the page's JSON-LD describes three of
  them rather than the results. premium_spotlight marks a result that is
  also paid.
• price needs a real asking price. A no-price listing's display price
  is its MSRP; salePrice means different things on different dealers and
  is not a column. A KBB 0 is null.
• A sold listing is gone, not 25 cars from the unrelated results page
  the site redirects to.
• The refusal page's reCAPTCHA files an unblock request with the site's
  staff, so that page is blocked and never solved.
• pyppeteer's page.authenticate does not work with current Chromium
  (Network.setRequestInterception was removed), and its bundled Chromium
  would not start here. The engine answers the proxy through the CDP
  Fetch domain, and --chromium-path drives Playwright's Chromium.
• No engine sets a User-Agent. The browser's own was served.

Tests

• 360-odd offline checks on fixtures cut from real pages by
  make_fixtures.py, which proves each parses identically to its original
  and refuses to write one that still carries the requesting address, a
  session id or a private seller's id.
• Ten planted faults, each confirmed to turn the suite red on the check
  that names it. Two did not on the first attempt, and the suite was
  changed rather than the prediction: a placement check that derived the
  placements from the rows (so it emptied exactly when the filter broke),
  and an end-to-end vehicle run whose sold car came first, where the loop's
  end-of-search check never looks.