prompt: route covariance-interpolator-rng-seed (#1450) → start_library

[Jammy2211]

Mind state for the PyAutoFit interpolator flake — the Class B release blocker
that killed the 2026.8.2.1 live release.

Paired with PyAutoLabs/PyAutoFit#1451 (fix) and PyAutoLabs/PyAutoFit#1450 (issue).

Normally this would land on main via prompt_sync_push; this session is
branch-scoped, hence a PR.

What changed

• draft/bug/autofit/covariance_interpolator_test_unseeded_rng.md →
  active/, header advanced to Status: issued with the issue URL.
• active.md — new covariance-interpolator-rng-seed entry.
• draft/feature/autofit/search_seed_reproducibility.md — new follow-up
  prompt (see below).

python3 scripts/lifecycle.py check → OK.

Why the registry entry is long

It records that the prompt's premise was wrong, which matters more than the
fix. The prompt attributed the flake to the unseeded np.random.random() in the
test body. Measurement says otherwise:

• Pre-fix failure rate 1.65% (33/2000, 95% CI 1.18–2.31%) — measured before
  touching anything, as the prompt's step 2 required.
• Seeding only np.random leaves std at 1.95 of 2.05; the test still fails.
  The dominant source is dynesty's rstate (Generator(PCG64(None)), OS
  entropy), with the stdlib random module used by initializer.py third.
• The tolerance is not mis-calibrated — a full search recovers
  25.0496 ± 0.0039, 500× tighter. The autouse limit_maxcall fixture is what
  turns the assertion into a coin flip.
• The sibling test_single_variable was flaking worse (3.6%) with no random
  call of its own.
• The LinAlgError guards are explicitly ruled out as a different root cause
  (rank-deficient fixture matrix, cond 1.3e17 — deterministic input).

The entry also records the known limitation the fix does not solve, so nobody
later reads "seeded" as "the assertion is now meaningful".

The follow-up prompt

search_seed_reproducibility.md captures the real gap the fix exposed:
PyAutoFit has no supported way to seed a search.
AbstractDynesty.search_kwargs is a closed dict, so rstate cannot be threaded
through even though dynesty accepts it — which is why the fix must monkeypatch
a third-party module from a test, and why users cannot reproduce a fit. It flags
the identifier/serialization trap (store the integer seed, not a Generator;
decide deliberately about __identifier_fields__).

Filed rather than folded in, to keep the release-unblocking change test-only.

🤖 Generated with Claude Code

https://claude.ai/code/session_016vdGab4C5rWumTCXkWHPyh

---

Generated by Claude Code