[edsonsims]

A connect-timeout callback delivered after cancellation can interrupt a ready query, fail after the socket has closed, or close a replacement connection. I added a generation and active-handle check to the shared timer helper so cancelled or replaced callbacks are inert. The active handle is cleared before invoking the handler, preserving rearming. Configured deadlines and public APIs are unchanged.

Refs #1270. The regression deliberately invokes captured callbacks after real local PostgreSQL startup, query, close and reconnect; it does not reproduce the reported Cloudflare timer-cancellation cause.

Validation with PostgreSQL 17.10:

• Four permanent cases: three fail on the original source; all four pass with the fix, including a stalled protocol-cancellation connection that must still time out.
• Full suites: 268 passes in ESM, generated CJS and Deno 2.1.4. ESM also passes on Node 24.19.0; the initial Node suites used 23.6.0.
• ESLint and the generated builds pass. The patch includes canonical source and tests, following the current source-only contribution convention; unrelated pre-existing generated-file drift is excluded.