refactor(adr-102): split serialization.py into a serialization/ package (P6d)

[aaronsb]

ADR-102 P6d — split the 1573-line serialization.py

Part of P6 (last ADR-102 spine phase). Follows P6a (#487) and P6b (#488).

api/lib/serialization.py was 1573 lines and repeatedly flagged by the >800-line
quality gate. It held four distinct responsibilities that map cleanly to modules,
so this splits it at the natural class seams into a package — zero caller changes.

Layout

Module	Contents	Lines
serialization/format.py	BackupFormat + KgBackupV2Reader + KG_BACKUP_FORMAT_VERSION (schema probe + pure read path)	196
serialization/exporter.py	DataExporter + _parse_nullable_int (export path)	703
serialization/importer.py	DataImporter + _execute_with_age_retry/_progress/_run_parallel + epoch primitives (clone/merge writer)	698
serialization/__init__.py	re-exports the public surface	19

Import-compat (the whole risk)

All existing from api.lib.serialization import ... callers keep working via the
__init__ re-export. Verified consumers: backup_archive, backup_streaming,
gexf_exporter, backup_integrity, restore_worker, restore_modes, plus 7 test
modules. The re-export covers every symbol any of them import
(DataExporter, KgBackupV2Reader, DataImporter, KG_BACKUP_FORMAT_VERSION; BackupFormat also exported).

No cross-class references exist between the three modules (verified by grep), so no
circular imports. format.py imports nothing from exporter/importer.

Incidental cleanup

Dropped the vestigial sys.path.insert(0, str(Path(__file__).parent.parent)) hack
and the now-unused sys / Path / Colors imports. The hack inserted api/ —
which doesn't even serve the absolute from api.app.lib.age_client import AGEClient
import (that resolves via the real PYTHONPATH). Confirmed harmless by the test run.

Why exporter/importer are still ~700 lines

Each is a single cohesive class (DataExporter / DataImporter). The three-class
boundary is the natural seam; splitting a single class further would be artificial.
Both are now under the 800-line priority threshold.

Tests

• ADR-102 + backup slice (kg_backup_v2 / reader / id_remap / restore_modes / restore_worker_epoch / kg_backup_v2_restore / epoch_reconciliation / backup_integrity / backup_streaming) — 93 passed, 1 skipped
• import api.app.main + public-surface hasattr assertions — OK
• All 4 modules ast.parse clean

🤖 Generated with Claude Code