Per-object timeout in the batch path

One object that never converges currently stalls the whole chunk. Fitting 108 short arcs in a single orbitfit call ran over an hour with no output; the same 108 fitted individually took under a minute of compute, because two of them grind indefinitely and nothing interrupts them.

At catalogue scale that is the difference between a run finishing and not. The pathological rate on short MPC arcs measured 1.9% at 7 days and 7.2% at 14 (#465). On 1.5M objects that is tens of thousands of fits that never return.

Suggested: a per-object wall-clock budget in the chunk loop, defaulting to something generous. On expiry, record the object with a distinct flag and move on. The result should say which objects were abandoned rather than silently omitting them.

This also makes a run resumable in a useful way — a chunk that hits the budget still produces output for everything else in it.

Start by locating the chunk loop behind the orbitfit call and reading how per-object results and failures are currently recorded. Define the timeout and abandoned-object status there, then verify that a timed-out object is reported while the remaining objects still produce output and the chunk can be resumed.