Set the solver defaults to what measured best, and end the drift

[adamjohnwright]

Depends on #30 (which makes the test suite run to completion); merge that first
or this branch's test changes are partly invisible.

The drift

The code defaults and docker-compose.dev.yml disagreed on four variables.
docker-compose.prod.yml sets no solver vars, so production ran the code
defaults while every benchmark ran through the dev container. specs/002
(2026-09-10) wrote its winning values into the code and never touched compose,
which still carried 2026-07-16 values.

The tests pin the code defaults — so the suite failed in the project's own dev
container, and since a failing top-level @testset aborts the file, 141 of
151 assertions silently never ran.

Attribution — one variable at a time

23,022 wide-curator cases, Release97, one catalog build. Each arm's container
env and catalog mount were verified, and its raw pred_ui checked to differ
from the baseline — a config change producing bit-identical output has not
applied.

variable moved	changed	net
DS_ASSEMBLY_LIMITING 1 → 0	613	−196
DS_AND_MODE hill_log → hill_sat	171	−90
DS_INHIBITOR_EPS 1e-3 → 1e-12	269	+11

Assembly-limiting is the driver, restoring the result already on file from
2026-07-16. The epsilon is not: its whole range is worth 11–14 cases against a
248-case gap.

Epsilon sweep

DS_INHIBITOR_EPS	correct	macro-F1	net
1e-3	19,278	0.8035	—
1e-4	19,292	0.8045	+14
1e-6	19,289	0.8041	+11
1e-12	19,289	0.8041	+11

1e-6 and 1e-12 are bit-identical — same 269 changes, same +109/−98. At or
below 1e-6 the epsilon has saturated into a pure divide-by-zero guard, exactly
as specs/006 FR-002 required and nobody had confirmed.

1e-4 scores 3 better in 23,022 and is not taken: it still changes results,
so it is still acting as a model parameter, and picking it would optimise a
guard constant against the evaluation set.

Result

DS_AND_MODE          = hill_log   (was hill_sat)
DS_ASSEMBLY_LIMITING = true       (was false)
DS_INHIBITOR_EPS     = 1e-12      (unchanged; compose corrected from 1e-3)
DS_HILL_SAT_EPS      = 1e-5       (unchanged; inert under hill_log)

Code and compose agree, all four assertion suites pass in the dev container, and
production computes what the benchmark measures.

The tension, recorded not buried

specs/002 R5 was not wrong that hill_sat implements the stated AND
intent: 10×10 reads 99.98 against hill_log's 74.06, and a lone node at UI 50
reads 50.00 against 41.43. hill_log compresses throughout the operating range,
which by that argument is a defect.

It also scores 90 cases better. So the compression is doing useful work that
"correct" multiplication does not, and nobody has explained why — a
plausible candidate is that it damps the multi-hop amplification behind
false-change calls, but that's a hypothesis.

This is therefore an empirical default, not a principled one, and
test_and_curves.jl names its testset accordingly instead of claiming the
defaults implement the design intent.

Full record: specs/009-solver-defaults/research.md.

🤖 Generated with Claude Code