feat: row-level parity diff tooling

[Hydaspex]

Final roadmap increment: row-level added/removed/changed reporting between record sets — usable for run-over-run weekly comparisons and, later, backend parity checks. Follows on from #5.

What's included

pipeline/diff.py

• diff_records(old, new) -> ParityReport — keys records on the same identity as normalise_records (region, provider, specialty, metric) and classifies every key as added / removed / changed / unchanged. Pure and deterministic (sorted-key output).
• RecordChange — kind, key, old, new, and changed_fields with field-level detail over source_url, average_wait_weeks, patients_seen_within_weeks, page_last_updated. A None → value transition counts as a change.
• ParityReport — is_identical, total_changes, summary().
• format_change — one human-readable line per change (+/-/~ with old → new values).

cli.py — new nhs-scraper-diff OLD NEW entry point: prints the summary and per-change lines; exit 0 when identical, 1 when changes found, so a scheduled workflow can gate on waiting-time movement.

io/csv_handler.py

• read_records_csv — strict header validation against the canonical COLUMNS (loud failure on foreign or moved schemas), None round-trip, tolerates legacy 8.0 float formatting, rejects non-integral week values.
• Bug fix found during sandbox verification: week columns mixed int/None previously became float64 and serialised as 8.0 in the CSV — now cast to pandas nullable Int64, so output is always plain integers. Regression test added (test_integer_columns_never_serialise_as_floats).

Tests (12 new diff + 6 new CSV) — identical/added/removed/changed/None-transition/mixed scenarios, format_change rendering, CLI end-to-end via capsys (exit codes + output), reader round-trip/None/legacy/header/non-integral cases.

Verification

Diff engine (all classifications, field detail, None transitions), the Int64 writer fix, and the reader were all executed in the sandbox — green. The float-serialisation bug was caught by that verification, which is exactly what it's for.

Design decisions to review

1. Same identity for diff and normalise — one definition of "same record" across the codebase; changing it changes both.
2. Exit-code contract (0 identical / 1 changes) — makes nhs-scraper-diff composable in CI without parsing output.
3. Strict header validation on read — a schema move or foreign file fails loudly rather than silently mis-parsing.