v0.1.1 — a moved shape is never an empty listing

Exit codes change for a page the parser cannot read. v0.1.0 treated a
listing record with no usable UniqueId or no Seller as advertising, so
a moved payload shape read as an EMPTY page: exit 4 on page 1 ("nothing
matched"), and a complete run with exit 0 when it happened on a later
page. It now ends as parser_found_nothing: exit 5 when nothing was read,
exit 6 (partial, rows kept) when earlier pages were. A pipeline that
treated exit 4 as "empty search" was being told something false.

Fixes from a third-party audit of v0.1.0, each reproduced before changing
anything and each pinned by a check that a planted fault turns red.

Fixed

• A record is sponsored only on a POSITIVE marker (MediaZeroKm or
  AdvertisementLink). Measured across every capture: 3 of 3 sponsored
  tiles carry both, 0 of 3,386 listings carry either. A listing that cannot
  be read is counted in the sidecar's new malformed_records and logged.
• Core columns that fall below their coverage floor are recorded in the
  sidecar (columns_below_floor), not only logged, and the canary fails on
  them and on any malformed record.
• Ad mode: the sidecar's query identifies the SET of adverts asked for
  (ads_sha256, order-independent) and lists them in ads_requested.
  v0.1.0 recorded only their number, so diff_runs.py compared two
  different lists of one length.
• diff_runs.py refuses two search runs of different depth
  (pages_requested): the extra pages read as listings the market added.

Changed

• A skipped canary run says so at the top of its summary page, and the
  README says the badge is green because the canary skips: this repository
  has no proxy secret set.

v0.1.0 — first release

First release. Two modes over webmotors.com.br's own JSON endpoints, three
browser engines over one shared fetch loop, and the 2Captcha Scraper API
client.

Added

• --mode search: a search page of cars (/api/search/car) or motorcycles
  (/api/search/bike), one row per listing. Any listing address the site
  builds is accepted with --url; --category, --condition, --state,
  --make and --model build one. The five orderings the site's own menu
  offers (--sort), and --per-page up to 100.
• --mode ad: one row per advert (/api/detail/{car,bike} plus
  /api/detail/averageprice), from --url or --ads-file (a search run's
  JSON output, or one address per line). An advert the site answers 404 for
  is recorded in the sidecar's ads_gone and does not fail the run.
• Row URLs built exactly as the site builds its advert links (660 of 660
  identical on the day), so a search's output feeds --mode ad directly.
• The currency, read once per run from a listing page's JSON-LD, since no
  endpoint states one.
• Refusals up front and after page 1 for what the API would otherwise answer
  with a WIDER search: an unknown state before anything is sent; a make or
  model the site did not apply, from its own FilterCustom echo.
• Sidecar fields total_results, pages_available, reachable_max,
  capped_by_site (the site's ~10,000-result cap), sponsored_skipped,
  currency and, in ad mode, ads_gone.
• tools/capture.py and make_fixtures.py: raw captures, then scrubbed,
  trimmed fixtures proven to parse like their originals.

Measured and built in

• CloudFront refuses datacentre addresses; the engines warn when run with no
  proxy and no --cdp-endpoint, and retry a refused page three times from a
  fresh browser, since a rotating residential gateway gives each one a new
  exit (5 of 12 accepted on a -region-br login, 9 of 12 on -region-us).
• PerimeterX refuses headless Chromium's HeadlessChrome user agent (0 of 3)
  and an override that drops the client hints (pyppeteer, 0 of 2). All three
  engines send a complete identity (page_flow.ua_override).
• pyppeteer answers a proxy's authentication over CDP's Fetch domain, since
  its own page.authenticate is dead on current Chromium, and connects to a
  Scraping Browser without ignoreHTTPSErrors, which hung it.
• Sponsored new-car tiles are skipped and do not shift position.

Not included, on purpose

• No captcha solver. The one challenge the site showed is PerimeterX's Press
  & Hold, which this repo does not implement, and no client it drives was
  shown it once the identity fixes were in. captcha_solver.py from the
  family core is not shipped rather than kept as dead code.
• No private seller's name, postal code, phone or description is written.