feat: add n_effective to DynestyDynamic; fix truncated runs under a finite iterations_per_full_update

[samlange04]

Summary

Discussion: https://github.com/orgs/PyAutoLabs/discussions/33

Two changes to DynestyDynamic, both motivated by making a dynamic run cheaper while still writing output on the fly.

1. n_effective argument. dynesty's dynamic sampler keeps adding batches until the estimated effective sample size reaches n_effective (dynesty 2.x default max(10000, ndim ** 2)). PyAutoFit had no way to set it. DynestyDynamic.__init__ now takes n_effective: Optional[int] = None, forwarded to run_nested only when not None, so default behaviour is unchanged. Not an identifier field. No yaml default, following refactor: replace search YAML config with explicit Python defaults #1202. DynestyStatic is left alone because dynesty deprecates n_effective on the static run_nested.

2. Fix: a finite iterations_per_full_update truncated a dynamic run. run_search_internal assumed dynesty's per-run call counter resets on every run_nested call. That is true for the static sampler but the dynamic sampler compares maxcall with its cumulative self.ncall, so the second chunk got a budget it had already spent, returned without sampling, and the "no new calls" criterion returned a cut-short baseline (ESS about 1, no batches). A baseline run cut by maxcall also cannot be continued (the next call reset()s it; resume=True refuses after RUN_DONE), and results.ncall undercounts the sampler's own counter (it omits batch live-point initialisation).

   The fix adds per-sampler hooks on AbstractDynesty (total_calls_from, maxcall_from, chunk_kwargs, chunk_is_finished) whose defaults are the existing static behaviour. DynestyDynamic overrides them: the first chunk of a chunked run completes the whole baseline (maxbatch=0, no budget), the batch phase is then chunked by cumulative budget, and a batch chunk is finished only if it stopped strictly inside it. Intermediate output for dynamic runs starts after the baseline. The run_search_internal docstring records the two samplers' counting conventions.

API Changes

• DynestyDynamic(n_effective=None): new optional keyword, default None (dynesty's own default applies).
• New overridable methods on AbstractDynesty: total_calls_from, maxcall_from, chunk_kwargs, chunk_is_finished. Defaults reproduce the previous static behaviour. DynestyDynamic.chunked property.
• Behavioural: DynestyDynamic with a finite iterations_per_full_update now returns a complete run; its first on-the-fly update happens after the baseline rather than mid-baseline.

Test Plan

• New tests in test_dynesty.py: n_effective is passed to run_nested only when set; the dynamic hooks give the baseline chunk an unbounded budget and maxbatch=0, cumulative budgets for batch chunks, and the strict-inside-budget finished criterion; the static defaults are unchanged.
• End to end, 3-parameter Gaussian, iterations_per_full_update=1500: before, ESS about 1 and no batches; after, baseline + 3 batch chunks + 1 no-op, ESS 3587 against 3944 for the same run in a single chunk, same logZ and parameters. n_effective=2000 stops batching earlier with the ESS floored at what the baseline gives. DynestyStatic unaffected (4 chunks of about 1500 calls, same result).
• test_autofit/non_linear/search/nest/: 44 passed, 13 skipped on macOS / Python 3.13 / dynesty 2.1.5, on the branch rebased onto main at 710f4b3.

🤖 Generated with Claude Code