Cache prepared statements per pooled connection

[adhikjoshi]

Laravel prepares every query fresh — one synchronous COM_STMT_PREPARE round trip plus parse work per query (PDO::prepare was 7.5% of gpulab PHP CPU, 4% frontend). An LRU cache (default 64) of PDOStatements per Connection, keyed by PDO instance + SQL, now sits in MySqlStringBindingConnection.

Covered paths: select, statement, affectingStatement, and MySqlConnection's own insert override (bypasses statement() to capture lastInsertId — every Eloquent create()). cursor()/selectResultSets() uncached by design.

Safety (all adversarially reviewed and test-pinned):

• Cache lives on the Connection (one coroutine at a time via the pool); setPdo/setReadPdo flush it (disconnect + reconnector).
• errno 1615 → evict + re-prepare once; errno 1461 (server stmt cap) → flush whole cache + retry once.
• select() drains cursors in a finally — on the unbuffered connections two apps run, this also heals the pre-existing abandoned-result wedge (2014 pool poisoning) that vendor code could not.
• Escape hatches OCTANE_MYSQL_STMT_CACHE / _SIZE, wired independently of OCTANE_MYSQL_STRING_BINDINGS (review catch: one hatch used to kill both features).
• Known semantic deltas documented in the class docblock (short-bindings HY093 now silent; StatementPrepared fires on writes/hits; last bindings pinned until reuse/evict).

Evidence: 190 tests green; 11 cache tests incl. lastInsertId-per-execution, stale-results, LRU, read/write segmentation, 1615/1461 heal, hatch independence; 6 mutation tests kill guards. E2E on real Octane+Swoole+MySQL: 200 → 2 prepares per 100 requests, 18-probe battery byte-identical in BOTH buffered and unbuffered modes, 10 concurrency rounds with zero errors, statements bounded and freed on connection close.

Rollout: staggered per app, watching Prepared_stmt_count (cap 65,536; steady state ~10.6k today).

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}