Widen live CI, and fix two proxy-failure defects it turned up

[jehrr]

The remaining half of the audit's "CI is narrow" finding, plus two defects
that looking for its "proxy failure semantics" test turned up.

The two defects

The proxy-failure classifier was declared three times. Each engine
carried its own copy of the Chromium marker tuple and inlined its own match
against it. They were byte-identical when this was written, which is the
state a drift starts from rather than a defence against one. It is now
page_flow.proxy_failure(), which takes the exception rather than
pre-extracted text so no caller can forget to stringify a driver's own
exception type.

And the responses had already drifted. A dead proxy and a timeout want
opposite answers — another try at the same exit, versus a different exit —
and pyppeteer rotated on any load failure:

playwright / selenium   if exit_failed and block_attempt < block_retries
pyppeteer               if load_failed and block_attempt < block_retries

So an ordinary network flap spent a --proxy-block-retries budget and
re-fetched the page through a fresh browser on another exit, while its twins
gave up and reported page_load_timeout. Three engines disagreeing about
what a timeout means, on the one invariant that names this exact confusion.

The CI half

weekly.yml covers what the daily canary does not: every mode (product,
reviews, a best-seller grid), a second and third marketplace, and the two
non-primary engines over two pages each. The assertions are chosen to be the
thing each one can be wrong about while everything else still looks right —
the currency per marketplace, page+position uniqueness per engine,
record_type and reviews that actually carry an id and a body.

The query is per-marketplace, because a German term sent to amazon.co.uk
still returns something, and a run that looks fine while exercising a query
nobody would type is worse than one that fails.

Dispatch-only, for the same reason the variation canary was: nobody has
watched these go green. The evidence that does exist is written into the
header along with what it does not cover, and the cron line sits
commented beneath it.

What is promoted

variation-canary moves to the daily schedule, on evidence: two dispatches
from GitHub runners with no proxy and no secret, both green.

Those two runs are also why its assertion stays a floor rather than the
measured figure. The same product reported 776 variants from a
workstation, then 744, then 745 from the runners ten minutes apart.
One product, three numbers, two hours — a canary pinned to any of them would
go red for the catalogue behaving normally.

Verification

• smoke_test.py: OK. .github/ci_checks.py --all: clean.
• Every new check controlled. Restoring pyppeteer's old rotation condition
  reddened both of its parity checks and named it.
• The inline-Python check added in the previous PR picked up all three of
  the new heredocs on its first run, which is that guard doing the job it
  was written for a day earlier.

🤖 Generated with Claude Code