fix(worker): give up on an instance only after failed starts in a row

[AlexCheema]

Problem

A worker asks for an instance to be deleted once it has created a runner for it EXO_MAX_INSTANCE_RETRIES (5) times. That limit is meant for an instance that can't start. But the count was only cleared when the instance was deleted, so it also counted runners that had started, served, and later died: a killed process, a peer's runner failing, a node restart. Over its life, a healthy instance used up its starts one restart at a time, and the next restart deleted it.

In a chaos test, a 2-node pipeline instance was deleted 9 minutes after it was placed, after two of its runners were killed. A pipeline instance recreates runners on every node when one dies, sometimes with a failed attempt along the way, so a few kills use up 5 starts.

Fix

Clear an instance's count when this node's runner for it is ready. The count now means "starts in a row that never got there".

Tests

• test_instance_retries.py drives the worker's real event loop:

  • a runner that starts and is killed 10 times never uses up the retries;
  • 5 failed starts in a row still add up to the limit;
  • another node's runner being ready doesn't count.

  The first test fails without the fix.

• On hardware: one Mac Studio (M3 Ultra) running Llama-3.2-1B. Six times, wait for its runner to be ready, then kill the runner process.

	Instance after kills 1–5	Kill 6	After
main	still running	instance already deleted	0 instances, 1 deletion request
this PR	still running	still running	1 instance, ready, 0 deletion requests

🤖 Generated with Claude Code