[plainprogrammer]

Closes #411

A direct fix, routed with the coordinator (no spec). A bulk-file line that isn't valid UTF-8 is now counted as one malformed line and the refresh goes on, as for any other bad line.

Cause

MTG::Scryfall::Source#entry_or_malformed called String#strip first. On a line that isn't valid UTF-8 it raises Encoding::CompatibilityError (or ArgumentError when the bad byte comes first, which was already rescued). The first wasn't rescued, so each_entry aborted mid-file and the whole refresh failed on one bad byte. Low impact (Scryfall's files should never hold such a line, and it failed safe), but one such line would have stopped every refresh until Scryfall fixed the file. Found in the implementation review of spec 085 (#364); the price source already has the fix.

Change

• app/models/mtg/scryfall/source.rb: the first line of entry_or_malformed returns Catalog::Sources::Malformed.new(external_key: nil, error: "the line is not valid UTF-8") when line.valid_encoding? is false, before any strip or parse. Same shape as MTG::Scryfall::PriceQuotes#each_card_record. No shared helper: Catalog::Sources::Malformed and Prices::Sources::Malformed are different types and each guard is one line, so the two stay as they are (the coordinator approved this).
• spec/models/mtg/scryfall/source_spec.rb: a gzip file with a 0xFF byte in a line before a good card, expecting a Malformed and then the good record.
• docs/architecture/catalog-operations.md: one sentence saying what counts as a malformed line now, A catalog refresh stops on one line with invalid UTF-8 instead of counting it as malformed #411 in Sources and a log row. Nothing else in the docs stated what counts as malformed.

Evidence

• Failing first: the new example failed with Encoding::CompatibilityError: invalid byte sequence in UTF-8 at source.rb:142 (String#strip); it passes with the guard. The whole source_spec.rb: 22 examples, 0 failures (also after merging main).
• bin/ci checks green; the pre-push hook ran the full bin/ci at 95b29867 (5048 examples) and passed.

Implementation review

Opus 5.5, read-only: SPEC-ALIGNED. It confirmed that valid_encoding? tests UTF-8 (the reader's lines come out in Encoding.default_external, as in the price source), that a Malformed with no key is counted like the existing {not json case (Catalog::Refresh#sync_entry), that the new example fails without the fix for the right reason, and that no wider rescue is justified. It asked for one comment correction (done: a bad first byte raises ArgumentError) and raised a possible gap, a lone surrogate escape making the refresh's digest encoding raise. I checked it with bin/rails runner in the test environment: json 3.1.0 rejects a lone \ud800 as JSON::ParserError, which the source already rescues, so there is no gap.

🤖 Generated with Claude Code

[brunnojob]

The guard is placed before both strip and JSON.parse, so malformed UTF-8 is classified as one malformed record instead of escaping the per-line parser path and aborting the whole refresh. It also keeps the existing blank-line behavior for valid input unchanged. The regression in this diff exercises the malformed-line count and continuation behavior. Static review; I did not run the Ruby suite.