fix(auth): define normalize input-open failure diagnostics

[stacknil]

Summary

Define normalize diagnostics for missing or unreadable inputs: retain exit 1 and empty stdout, and report the input path, error type and message on stderr without a Python traceback. Addresses #7.

Why

The README and tests define resilient per-line parsing but leave startup failure diagnostics unspecified. This closes a CONTRACT_GAP without claiming a violation of a prior no-traceback guarantee.

The baseline evidence and classification were posted before the signed commit and this PR.

Behavior change

• Catch OSError only during automatic input detection and input opening; preserve existing output files.
• Use str(exc) when OSError.strerror is None.
• Keep completed-batch exit 0, per-line diagnostics and --error-output JSONL unchanged. Startup failures do not produce per-record error rows.

Validation

• Missing-input and permission regressions previously demonstrated red-to-green behavior; the additional missing-strerror regression also failed before its fix and passed afterward.
• Focused CLI: 14 passed; complete auth suite: 24 passed, on Windows Python 3.12.14 and WSL Python 3.12.3.
• Applied the final patch to a clean auth-lab archive of 9b8ae31b7ea89eddf7bb4f8b1ce6c131e781b627; applied files match the tested candidate. WSL tests ran on that reconstructed candidate.
• Syntax compilation and git diff --check pass. Local Python 3.11 was unavailable; GitHub Actions on Python 3.11 passed all four mini-labs: auth 24, socket 8, permission 6, process 30.

Compatibility

Exit codes, record schemas, source detection and the normalize/filter/summary workflow remain unchanged. Consumers parsing startup traceback text will see the intentional diagnostic change. Output failures remain outside the new catches; no dependency is added.