Skip the redundant checkout round-trips the pool paid on every borrow

[adhikjoshi]

Problem

Datadog spans from production showed the pool's own overhead as the top SQL by count and total time: every checkout paid a SELECT 1 ping plus two session SETs, and every release paid the SETs again — 3–5 round-trips at ~10–20ms each from Cloud Run to the DB host. That is ~30–50ms of latency tax on every DB-using request in every app, and 162M SET statements on the MySQL server in 21 days.

Change

release() fully resets every connection before re-pooling and closes any whose reset fails, so pooled connections are clean by invariant. Checkout now:

• runs a local dirty check (transaction counter + raw-PDO flag, no SQL, no lazy connect) and only pays the full reset for a genuinely dirty connection;
• pings only connections idle past ping_after_idle (default 30s); freshly pooled/created connections skip it — a connection that died while idle is still recovered by Laravel's lost-connection retry through the pool's identity-preserving reconnector;
• fresh connections get their session normalized once at creation (same SETs as the reset), so fresh and recycled connections are indistinguishable even when the server's global isolation/autocommit defaults differ.

Verification

• E2E (real Octane + MySQL): 50 requests → session SETs 104 → 51 (exactly one reset per borrow, release-side), checkout pings gone.
• 171 package tests green; new behaviors mutation-tested (dirty-check removed, ping removed, idle-order swapped, normalization dropped → each test goes red).
• Adversarial review round completed; its findings (session-normalization gap, lazy-PDO resolution in the dirty check, doc claims) fixed and covered.

---

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