[kvlonge]

Why

A deadline cancellation marked the interrupted connection by id(connection) in a process-wide set. Only a recovering connection pool's health check cleared marks, so a connection that never returned to such a pool left its mark behind. Once that connection was collected, a later connection allocated at the same address was health-checked as if it had been interrupted, running an extra probe statement and publishing its statement events. This made the scenario capture pipeline's connection-lifecycle unit test fail intermittently in CI. Separately, the 1000-model build memory guard sat within 0.2-0.5% of its cap on main, so unrelated branches failed it at random.

Changes

• InterruptedConnections keeps a weak reference to each marked connection. Objects that can't be weakly referenced get a strong one. A check reports only when the mark still refers to the same connection, and each check still clears the mark.
• The 1000-model build memory guard goes from 1,728 to 1,760 MiB: main measures 1,778-1,808 MB on CI runners (about +/-15 MB noise) against the old 1,812 MB cap.
• No change for connections that really were interrupted: they are still checked once and replaced when unusable.
• Unit tests cover single reporting with and without weak references, and a new connection reusing a collected interrupted connection's identity. That last test fails without the fix.

Verification

• Forcing a mark onto the capture pipeline's connection reproduces the CI failure: two extra statement events.
• Focused execution-limit, scheduling, pipeline, load and executor integration tests pass (517).
• ruff, ty and fensu are clean, make check-ci passes, and the hygiene scan is clean.