Report exit 5, not 4, when the content was never obtained

[jehrr]

A run that never obtained its pages used to report exit 4 — the code
this family documents as "ran fine, found nothing". So a dead proxy and an
empty catalogue were one value to anything branching on the exit code, which
is what this family says exit codes are for.

That much was already known. What a family-wide measurement on 2026-09-21
found is that three different repos had fixed it three different ways,
none aware of the others. Measured by importing each repo's output_writer
and CALLING finish_run([], …, stop_reason="page_load_timeout"):

answer	repos
exit 4	20 — unfixed
exit 5	2 — farfetch, amazon
exit 6	7 — fixed, with written reasoning, using EXIT_PARTIAL
neither	1 — stockx, not built on this core

This unifies all 30 on exit 5, and takes the better half from each camp.

The rule comes from the exit-6 camp

They keyed on if not complete. The exit-5 camp keyed on a named list of
stop reasons. A list cannot cover a reason nobody has added to it yet, so
a new stop_reason falls silently through to "the catalogue is empty" — the
exact defect all three camps set out to fix. So the list was the hole, and
the repos that shipped the fix first are the ones being simplified here.

The code comes from the exit-5 camp

EXIT_PARTIAL (6) means "some rows were gathered and the output is
incomplete". A run holding nothing writes no output at all, so a consumer
that reads the file on a 6 finds either nothing or the previous run's
good data — which save deliberately does not overwrite. Exit 5 promises no
file, and this family's contract already reserves 5 for a transport failure
(scraper_api_client has used it for a remote API error since it was
written). No new code, and no per-repo table for a caller driving more than
one of these scrapers.

A run that gathered rows and THEN failed is still exit 6, unchanged.

Verified by calling, not by reading

Every repo's finish_run was invoked after the patch and required to return
5 before the change was kept, and each repo's own suite had to stay green.
Where a repo's suite pinned the old answer by name, that check was
rewritten in the same commit — a test asserting exit 6 is that repo
stating its position, and changing the code while leaving the test would
make the suite red and read like a regression rather than a decision.

🤖 Generated with Claude Code