Re-bless the scenario gateway and diff its numbers

[dfalster]

scenario_baseline.rds matched no build since #583, so
PLANT_RUN_SCENARIOS=1 failed on an unmodified develop and every PR
that ran it inherited a failure it had not caused.

Bisected: phylloptim 0.2.0 -> 0.6.0 (in #633) moves R0 by 8x-32x on five
scenarios and takes S01 across R0 = 1; #619 moves five again on top. Not
#590. Everything else since is inert. Re-blessed on the develop tip as
its own commit, table in NEWS. The gateway now also diffs
offspring_production (SCENARIO_TOL, default 1e-3) and prints the
whole table on failure.

Closes #639.

[dfalster]

Bisection

One build per commit, each against the phylloptim/odelia sha that commit pins (installed from a git archive of the pinned sha into a scratch library, so no sibling working tree leaks in). macOS arm64, R 4.6.1, -O2 via pkgbuild::compile_dll(debug = FALSE), max_patch_lifetime = 100, SCENARIO_WORKERS = 8. Worker count does not enter the numbers — the scenarios are independent runs.

offspring_production:

scenario	baseline	#583 133499b8	#606 39c0c657	#633 b374ce3b	#619 745dd600	develop 60708313
S01	1.488721	1.488721	1.488721	0.2725883	0.4797763	0.4797763
S02	9.963559e-2	9.963559e-2	9.963559e-2	2.784370e-3	1.153315e-2	1.153315e-2
S03	5.498561e-12	5.498561e-12	5.498561e-12	1.745197e-10	1.665416e-10	1.665416e-10
S04	4.745912e-13	4.745912e-13	4.745912e-13	1.357792e-11	1.317815e-11	1.317815e-11
S05	5.888102e-13	5.888102e-13	5.888102e-13	5.052547e-12	2.470008e-11	2.470008e-11
S06	3.261466e-13	3.261466e-13	3.261466e-13	5.082346e-12	6.570073e-12	6.570073e-12
S07	1415.200	1415.200	1415.200	837.6889	810.7050	810.7050
S08	3.230258e-14	3.230258e-14	3.230258e-14	2.711131e-13	2.025578e-14	2.025578e-14

Step ratios:

scenario	-> #633	-> #619	net
S01	0.183	1.76	0.322
S02	0.0280	4.14	0.116
S03	31.74	0.954	30.29
S04	28.61	0.971	27.77
S05	8.581	4.889	41.95
S06	15.58	1.293	20.14
S07	0.592	0.968	0.573
S08	8.393	0.0747	0.627

persists changes once, at #633: S01 TRUE -> FALSE. Nothing else moves on either flag anywhere in the range.

b3472d43 (#608) was measured rather than b374ce3b (#633) — the two commits between them, #607 and #608, touch only NEWS.md, test-leaf.r and notes/, so they compile to the same library. The #633 column is that measurement.

Which half of #633

Rebuilt b374ce3b against phylloptim 0.6.0 with odelia held at 0.2.1 (the LinkingTo floor relaxed for the probe; nothing else touched). All eight scenarios identical to the odelia-0.3.1 build, so odelia 0.2.1 -> 0.3.1 is inert on the gateway and the whole move is phylloptim 0.2.0 -> 0.6.0.

That fits odelia's two behavioural changes in the range: stop_domain step rejection (odelia#56) had no declared domain to act on until #619 added Patch::ode_state_valid, and rejecting a non-finite error estimate (odelia#54) only bites a step that was already diverging.

Why #590 looked like the cause and is not

• 53d0a200 (Prototype of age integration #590) is an ancestor of the commit the baseline was blessed at.
• The baseline's own metadata records node_density_in_birth_date = TRUE — the recorded numbers were already in birth date.
• The magnitude coincidence is real but is a coincidence: Prototype of age integration #590's documented 2.4x-47x and Migrate to phylloptim 0.6.0 and odelia 0.3.1 #633's 8x-32x are both "a small perturbation to a water-limited run, read through an exponentially small survivorship integral".

Provenance of the old baseline

git_commit = 99e44db6, git_dirty = TRUE. It could not have been produced by that commit as committed: there, run_scenarios() still defaulted to control() (height) and scenario_run_metadata() had no node_density_in_birth_date field to write. The tree that produced it is the one that landed as 133499b8, which is why that is the control and reproduces it exactly.

Verification of the new test

Against the stale baseline it now reports, in full:

Scenario outcomes changed vs baseline (re-bless via `make bless-scenarios` if intended): persists S01: TRUE -> FALSE | offspring_production moved on 8/8 scenario(s) (tol 0.001)
id         baseline        current        rel
S01        1.488721      0.4797763     -0.678 *
S02      0.09963559     0.01153315     -0.884 *
S03    5.498561e-12   1.665416e-10       29.3 *
S04    4.745912e-13   1.317815e-11       26.8 *
S05    5.888102e-13   2.470008e-11       40.9 *
S06    3.261466e-13   6.570073e-12       19.1 *
S07          1415.2        810.705     -0.427 *
S08    3.230258e-14   2.025578e-14     -0.373 *

which reproduces the table in #639 exactly. Against the re-blessed baseline it passes.

Tolerance

SCENARIO_TOL defaults to 1e-3. It is not measured against cross-platform drift — the gateway is opt-in and has only ever been run on macOS arm64 — so it is set from the other end: small enough to catch the smallest move anyone has wanted to see (#633's +0.36% on the reference run), large enough to sit well clear of floating-point reordering. If a platform proves noisier, raise it there rather than widening the default; the two classifications are the part that is meant to be platform-stable.