Write the result of a calibration that ends before the search finishes #359

Description

When the search ends early, the runner consolidates evaluations.csv and re-raises (runner.py#L567-L582): on Ctrl+C, on a worker death (BrokenProcessPool) and on any other failure. result.json, stations.csv and <config>-calibrated.json are not written, and the documentation says so ("An interrupted run": what needs the finished search is not written). None of the three needs the finished search or a model run, though: the best record already carries the parameters, the objective, the NSE and the station_metrics of every station (_worker.py#L336-L384); only best_<variable>.csv runs the model again. A user who interrupts a calibration after hours, or whose worker was killed for memory, reconstructs the best candidate from the CSV by hand.

With the MODFLOW coupling of #357 a dead worker stops being exceptional: with dis.nstp > 1, a candidate whose solver fails ends its worker process, and with it the whole calibration.

Proposed solution

• On any early end after the first recorded evaluation, after consolidating the table, write result.json, stations.csv and <config>-calibrated.json from the best record. No model run on the way out, since the pool may be broken: best_<variable>.csv is not written and artifacts lists only what was.
• result.json gains a termination field: finished for a search that ran to SciPy's end, interrupted (KeyboardInterrupt), worker_died (BrokenProcessPool) and error (any other exception, described in message). success keeps the optimizer's meaning and is false on an early end; nit is the last generation the callback saw and nfev the number of records. This field is the vocabulary the stopping rules extend.
• A finished search writes result.json before the re-run of the best candidate and rewrites it with best_series once the re-run succeeded, so that a failed re-run leaves a complete summary minus that file instead of no summary (today "result.json not written at all", "Artifacts of a run").
• The exit codes and the messages of the command stay (2 on an interruption, 1 on a dead worker); the message names result.json next to the table.

Tests that prove it: the existing interrupted and crashed search tests (_record_then_raise(KeyboardInterrupt()), BrokenProcessPool) assert the three files and the termination value; a search that recorded nothing leaves no summary; a failed re-run of the best candidate leaves result.json without best_series.

Documentation: "An interrupted run", the result.json field list, the note on the failed re-run; changelog.

Alternative solutions

• Run the best candidate once more on the way out to write best_<variable>.csv. Rejected: after a BrokenProcessPool the pool cannot be trusted, and a user who pressed Ctrl+C does not want one more model run.

Additional context

Part of the calibrator robustness work assessed on 2026-09-24. The termination field is born here; #360 (stopping rules) and #361 (resume) build on it. Related: #358, #362. Whoever lands after #357 rebases runner.py.