[amanusk]

Summary

Closes #92.

Per WS event on a viewed contract, the Calls tab did O(n) work and a full SQLite rewrite of the address's rows: a linear dedup scan, push + sort of the whole list, a linear lookup per enriched row, then a clone of the entire list into PersistAddressCalls, which deleted and re-inserted every row. Gap rows were located with a linear position() per frame.

Changes

• Indexed, sorted Calls list (AddressInfoState): a tx_hash → block map gives O(1) dedup; insert_call uses partition_point into the block-descending order (same placement as the old stable sort of a push); find_call_mut goes hash → block → binary search into that block's few rows. Gap render positions for both tabs binary-search the sorted key (block for calls, nonce for txs) instead of position().
• Upsert-by-hash persistence: address_calls is re-keyed by (address, tx_hash) with a block_number column. save_address_calls now upserts only the rows it is given. The reducer sends just the rows a batch created or changed.
• Shape-keyed migration: detected via pragma_table_info, not user_version, so it runs on any existing DB regardless of which version bumps it has seen. Old rows are copied in one transaction with json_extract, invalid JSON skipped.
• One merge helper (merge_call_into) replaces the duplicated field-upgrade logic in deduplicate_contract_calls and the enrichment reducer. Enrichment now also fills inner_targets on streamed rows, so the Prv column no longer needs a restart.
• Network-side saves all go through persist_calls_merged (field-merge with cached rows, then upsert the touched hashes). Two non-Dune call builders previously never saved their rows and relied on the reducer's full-list write; they persist now. fetch_more_address_txs no longer overwrites the address's other rows.
• Stubs persist on insert, so a failed enrichment still leaves a row for the load-time repair pass.

Verification

• cargo test: 296 lib tests pass, including four new ones (insert order + dedup, hash lookup, gap positions for both tabs, migration + upsert against an old-shape table). cargo clippy --all-targets adds no warnings on the touched lines. Release build passes.
• Smoke run with a scratch config whose cache.db held a copy of a real address_calls table in the old shape, then streaming the public STRK token contract:

Check	Result
Re-key migration, 259,659 rows	1.46 s, one transaction
Writes during 90 s of streaming	133 upserts of 1 row + one 355-row Dune batch
Final run, in-memory vs DB Calls count	3194 vs 3194
Search paste visible while streaming	12 to 38 ms
50k-row microbench (call_insert_and_lookup_timing, debug)	insert 2.2 ms → 137 µs, lookup 1.4 µs

Notes

• First start after upgrade migrates the cache once (about 1.5 s per 260k rows). Nothing else changes for the user.
• Behaviour deltas to be aware of: enrichment joins endpoint names instead of only filling an empty one (same rule dedup already used), and deduplicate_contract_calls now also replaces a "?" status with a known one.

🤖 Generated with Claude Code