Bound large database comparisons and structure JSON failures

[Widthdom]

Summary

• Short-circuit comparisons of the same SQLite snapshot before row enumeration, including safe URI, symlink, and hard-link aliases.
• Detect separate sidecar-free byte-identical database copies with fixed-size buffers while retaining existing row and per-row byte budgets for distinct inputs.
• Return typed, parseable import --check --json budget failures and actionable human diagnostics.

Root cause

Large identical databases always entered the bounded streaming row comparison, so an identical input could exhaust the one-million-row safety budget. The import boundary also treated comparison-budget exhaustion as an unclassified failure. Filesystem identity alone was insufficient for aliases because immutable-mode and path-specific SQLite sidecars can expose different snapshots.

Validation

• Release solution build: 0 warnings, 0 errors.
• Focused diff/import suite after the final review fixes: 111 passed on net8.0 and 111 passed on net9.0.
• Full solution suite: net8.0 passed 11,182 tests with 7 skipped; net9.0 had one unrelated timing assertion failure, and its immediate focused retry passed on net8.0 and net9.0 (Stabilize candidate-ordered serial recovery timing assertion under full-suite load #5063).
• The repository's current large database now returns identical/exit 0 for a same-path summary diff.
• Changelog fragment validation: 20 fragments validated.
• Codex adversarial review completed in two rounds; all reported WAL/alias and recovery-guidance findings were addressed.

Base

This branch is intentionally based on d444cfdb2c660425afda504b9998977cdf15c89e. Newer origin/main commits were not incorporated per maintainer instruction.

Documentation and changelog

• Updated the English and Japanese sections of DEVELOPER_GUIDE.md and TESTING_GUIDE.md.
• Added changelog.d/unreleased/5052.fixed.md.

Follow-up

• Stabilize candidate-ordered serial recovery timing assertion under full-suite load #5063

Fixes #5052