feat: seedable multi-start draws, alive-versus-step curve, and optional momentum reset

[Jammy2211]

Three MultiStartGradient additions the prior-support validation campaign
(autolens_profiling#131) could not be run without. All are additive and
default-off/None, so existing fits are bit-identical.

1. seed — the multi-start draws were never seedable

The search's two random draws — the broad starting points and the resurrection
redraw — were hardcoded to default_rng(0) / default_rng(1). Seeding
random/numpy reaches only the initializer, so every run of a model drew the
same starting population: a "multi-seed" study was silently a single-seed one.

The two streams are derived through SeedSequence rather than by offsetting the
seed. seed + stream would make seed 0's resurrection stream identical to seed
1's starting stream, so nominally independent seeds would share draws and
resurrection would replay the starting population.

seed=None keeps the historical fixed seeds exactly.

This matters more than it sounds. With the argument in place, identical
settings swing the answer 171,272 nats between seeds 0 and 1 on
imaging/mge hst — invisible before.

2. alive_history — the budget-independent survival curve

The number of living lanes per step, written to search_internal.

The existing lane counters are survival integrals: a dead lane keeps adding
to them every subsequent step, so the same death curve reads ~60% at 150 steps
and ~75% at 300, and two runs at different budgets cannot be compared on the
scalar at all. The curve is the budget-independent quantity, and until now it
existed only in the progress log at iterations_per_log cadence — visible to a
human reading stdout, unavailable to any analysis.

3. reset_momentum_on_clip — optional, and measured NOT to pay off

Zeroes the optimizer moments wherever the clipper (#1477) just projected a
coordinate back onto its prior box. Default False.

Projection alone leaves a lane holding the exact velocity that carried it out of
the box, so the next step drives it into the same wall and it is re-projected
onto the same bound indefinitely — counted alive, permanently pinned, still
paying a full likelihood-and-gradient evaluation every step. The reset is
per-coordinate: a lane clipped in one parameter keeps its momentum in the others.

Moment fields are targeted by name, not by shape. Shape matching is actively
wrong here: Prodigy's params0 and grad_sum carry the same
(n_starts, n_params) shape as the moments, and params0 anchors its
learning-rate estimate — zeroing it would corrupt the step size for the rest of
the run rather than resetting momentum. A regression test pins that.

The measurement says don't use it (autolens_profiling#131, imaging/mge hst,
16x3000, fp64): same converged answer, but deaths 2 → 2523, one more lane
pinned, +39% wall on seed 0, and on seed 1 it gives back nearly all of plain
clipping's gain (-120880.6 → -137783.6). It ships default-off so the measurement
is reproducible, not because it is recommended.

Tests

test/non_linear/search/mle/test_multi_start_gradient.py — same-seed
reproducibility, cross-seed divergence, stream independence, alive_history
length/monotonicity, per-coordinate reset, and the params0/grad_sum
must-not-be-zeroed regression.

Follow-up

The campaign's diagnosis is that lanes reach the walls because box widths span
40x while the search steps in physical space with one global step scale — the
cause-side fix (per-parameter step scaling) is queued as a separate task, and
#1481 tracks which other searches need prior-support handling at all.

🤖 Generated with Claude Code