Importer perf: batch M2M name->pk lookup in link prepare

[ajslater]

Summary

Implements tasks/importer-perf/01-link-batching.md (planning PR #627). The headline impact-per-LOC item from the importer perf plan.

link_prepare_m2m_links previously fired one SELECT per (comic, M2M field) pair — for a 600k-comic import, ~6.6M small SELECTs in this phase alone, dominated by round-trip overhead. Restructured into three phases:

1. Collect — walk LINK_M2MS once and group every key tuple by M2M field name.
2. Resolve — one batched SELECT per field builds a {key_tuple: pk} dict.
   • Single-column models (named M2Ms, folders): IN clauses batched at IMPORTER_LINK_FK_BATCH_SIZE (30000) to stay under SQLite's 32766-variable cap.
   • Multi-column models (Credit, StoryArcNumber, Identifier): Q-OR chains batched at 500 to keep SQLite's planner well-behaved.
3. Stitch — per-comic loop becomes pure dict lookups, no SQL.

Query count drops from ~11N+1 to ~K+1 (K ≈ number of distinct M2M models, ~12). FTS side-effects preserved exactly — complex/folder fields still flow through _add_complex_link_to_fts, named M2Ms still hit add_links_to_fts with the same value shape.

Drops four now-dead helpers (_get_link_folders_filter, _get_link_complex_model_filter, _link_prepare_complex_m2ms, _link_prepare_named_m2ms).

Test plan

• make fix clean
• make lint-python clean (0 errors, 0 warnings)
• pytest tests/importer/ — 3 passed
• pytest tests/test_search_fts.py tests/importer/ — 7 passed (FTS side-effects intact)
• Field check on a real fixture: import a 1k-comic library twice (cold + warm), assert resulting Comic, M2M through-row counts match the pre-PR baseline byte-for-byte
• Wall-clock measurement on a real-scale fixture once one is available

🤖 Generated with Claude Code