[scratch] ISBN corpus validation harness (do not merge)

[raphi011]

⚠️ Scratch / review-only — do not merge

A throwaway local validation harness for metadata.isbns (from PR #2). It runs the full public pipeline (AssetRetriever.retrieve → PublicationOpener.open → publication.metadata.isbns) over a real corpus of EPUBs in ~/ebooks and reports correctness flags. It reads an absolute host path and skips when the corpus is absent, so it's a no-op in CI. Based on feat/epub-isbn-metadata (the harness calls metadata.isbns, which isn't on develop yet).

What it checks

• Summary: scanned / opened / open-failures, and ISBN-count distribution (0 / 1 / 2 / 3+).
• Flag A — false negative: the package OPF text contains isbn (case-insensitive) but metadata.isbns is empty → a marker we don't detect.
• Flag B — false positive: a returned value isn't a plausible ISBN (after cleaning, length ≠ 10/13, or chars outside [0-9X]).

It's a report, not a pass/fail gate (the flags need human review); the only assertion is that ≥1 book opened.

Run

xcodebuild test -scheme Readium-Package \
  -destination 'platform=iOS Simulator,name=iPhone 17' \
  -only-testing:ReadiumStreamerTests/ISBNCorpusValidationTests

Results (329-book corpus, 10.6s)

• 329 opened, 0 failures, 289 (88%) with ≥1 ISBN. Distribution: 40 / 280 / 9 / 0.
• Detection correct for every targeted marker (opf:scheme, value prefix, ONIX identifier-type refine) — no parser bugs.
• 8 misses (Flag A) + 2 p…-prefixed values (Flag B), all from nonstandard markers (ISBN only in the id attribute / dc:source / unmarked; Penguin p-prefixed values returned verbatim). Decision: leave as-is — these are a deliberate non-goal for the minimal parser.

[raphi011]

Closing — scratch validation harness was opened by mistake; not intended as a PR. Branch kept locally for re-runs.