Phase 15: re-measure parity under the finished epic (#351)

[neuromechanist]

Summary

Re-measures every user-facing parity figure with the finished epic #324 code (09b8247 + #353; merged with Phase 17, bb22df6, before review), against the pinned v0.3.3 native binary, and updates the documents that quote them: paper.md (Table 1, Validation, figure), docs/guides/validation.md, docs/guides/amica-differences.md, docs/guides/backends.md, ADR 0003 (new dated section), AGENTS.md, .context/progress_summary.md, benchmarks/README_dimsweep.md, the changelog, and the regenerated docs/assets/figures/multimodel-ensemble.png. Scripts and raw outputs are under .context/issue-351/; the findings record (findings.md) is committed by the lead.

No library code changes. benchmarks/reproduce_table1.py now pins every reference setting explicitly (REFERENCE_SETTINGS), and its ensemble's lrate, so Phase 17's default changes cannot move the Table 1 protocol. Before Phase 17 the pinned and unpinned calls wrote byte-identical input.param files at all 8 call sites (.context/issue-351/pin_check.py). After the Phase 17 merge each pinned call writes the same settings as before, in Phase 17's key order and with do_approx_sphere 1 added, the binary's compiled value (now pinned in REFERENCE_SETTINGS as well); the binary gives byte-identical output files for both forms of the file (pin_run_check.py, bundled sample, three configurations). Without the pins, 15 settings of the native calls would have moved to Phase 17's defaults.

Part of #324, closes #351 on the epic merge.

Old vs new

Figure	Before the epic	Now	Record
Harness (100 it, independent starts): dLL torch / NumPy / MLX	2.9e-5 / 2.9e-5 / 3.2e-5	2.71e-4 / 2.72e-4 / 2.76e-4 (reference's own seed-to-seed LL sd 2.6e-4)	raw/harness_run1, raw/harness_gap_head.json
Harness: mean / min corr, Amari	0.9992 / 0.9935, 0.0037	0.9991 / 0.9918, 0.0038	raw/harness_run1
Harness from a shared start: dLL, corr, Amari	2.35e-4, 0.99999, 5.8e-4 (e38aa11)	1.6e-6, 0.99999993, 3.9e-5	raw/harness_gap_*.json
Harness runtimes Fortran / torch / NumPy / MLX (s)	10.1 / 17.9 / 32.5 / 3.2 (same session)	10.0 / 18.1 / 32.6 / 3.0	raw/harness_*
200-it amicaout fixture: dLL, corr, Amari	2.2-2.3e-4, 0.9973, 6.3e-3 (pre Phase 11)	1.3-1.4e-4, 0.9983, 4.8e-3	raw/fixture_parity_head.txt
MLX f32 vs torch f64, same start, 100 it	not characterized	dLL 5.0e-6, corr 0.99999991, Amari 5.0e-5	raw/precision_agreement.json
Table 1 Amari (bundled, 5 pairs)	0.006	0.011 (one reference run in a lower-LL basin); 0.004 over 50 pairs (reference vs itself 0.005)	raw/table1_bundled, raw/bundled_basins
Score / sufficient statistics	~1e-15	8.9e-16 / 1.8e-15; 4.5e-13 abs (3e-16 relative)	raw/table1_bundled
Multi corr cross; within-Fortran	0.65; 0.64	0.632; 0.626 (perm p 0.88)	raw/table1_bundled, multimodel_summary.json
Multi Amari cross; within-Fortran	0.163; 0.174 (p > 0.999)	0.172; 0.166 (p 0.051)	same
Multi ensemble LL Fortran; pamica	-3.3539; -3.3629 (KS p 6e-5)	-3.3543; -3.3541 (KS p 0.83)	same, raw/multimodel (pre-epic refit -3.3627)
keep_best: sd ratio at 100 it, restores	12.7x -> 2.0x	1.0x with or without; 1 restore in 20 fits (300 it only, gain 4.2e-6)	raw/keep_best
Mean LL gap at 100 / 200 / 300 it	-0.009 / 0 / +0.002	+8.1e-4 / +2.1e-4 / +1.0e-4	raw/keep_best
Cross-backend 25-it LL (dimsweep)	max pairwise ~0.003	1e-5 at 32/48 ch; up to 1.1e-3 at 70 ch (k 6, round-off sensitivity 4.4e-4)	raw/dimsweep
Newton, independent seeds (70 ch)	seed 42: 10-11/70 below 0.9; others ~0.995	seed 42: 8/70 vs F1; others 0.995-0.996; reference vs itself 0.985 (2/70)	raw/newton
Newton, same start	0.9974 (min 0.9475)	0.9999998 (min 0.9999964), dLL 7.6e-8	raw/newton
share_comps e2e (300 it, thr 0.95)	3 merges, LL -3.3416	1 merge (	cos
Early scans at it 8 / 20 (0.95, 0.99)	32, 24 / 24, 5	30, 17 / 22, 5	raw/share_comps
Reference-formula scan on its own state	32, 32, 30	32, 32, 29	raw/share_comps
Table 1 external LL (on002718, 5 seeds)	gap ~0.0003 at -3.6993	Fortran -3.699346, pamica -3.699341, gap 5.6e-6	raw/table1_external
Table 1 external correlation	0.998 (Fortran vs Fortran ~0.999)	0.9996, min component 0.960 (Fortran vs Fortran 0.998); Amari 0.0013 (0.0025)	raw/table1_external

Notes

• The harness dLL and the 70-channel dimsweep spread are numerically larger than before; both were reported to the lead with evidence (start-to-start spread and round-off amplification) before the docs were edited.
• The old ensemble LL gap came partly from the pre-Apply the LL-decrease response before the update #339 min_dll check (7 of 20 pre-epic fits stopped early) and partly from that code's update rule; the "convergence speed" explanation is removed.
• The AMICATorchNG diverges from Fortran on weak components at 2000 iters, despite Newton making the optimum sharply unique #145 reference runs came from the gfortran amica15_linux build, which draws a new start on each unseeded run (checked on hallu), so both reference-vs-reference figures are reported side by side.
• Pre-epic diagnostics use e38aa11 (Phase 6), loaded as a separate tree; the orientation of the seeded state was checked by the first-iteration LL (9e-16).

Test plan

• uv run ruff check ., uv run ruff format --check ., uv run ty check ., typos
• uv run pytest -n 8 --no-cov with CI=1: 1955 passed, 40 skipped, 3 xfailed before the Phase 17 merge; 1974 passed, 40 skipped, 3 xfailed after it
• AMICA_RUN_FORTRAN=1 on test_fortran_param_forwarding.py and the two early-merge oracles: 37 passed, before and after the Phase 17 merge
• uv run --extra docs mkdocs build --strict
• pin_check.py: byte-identical input.param at all 8 native call sites before Phase 17, with a negative control; after it, the same settings at all 8 (--against the saved files), plus do_approx_sphere 1
• pin_run_check.py: the binary's outputs are byte-identical for the pre- and post-Phase-17 forms of the pinned files (wall-clock times removed from out.txt)
• External tier numbers integrated (lead's run at 09b8247, 5/5 seeds, 26185 s)
• Epic merged, including Phase 16 (3fbe1ab) and Phase 17 (bb22df6); one conflict in the Newton section of validation.md, resolved with Phase 17's wording of the do_newton defaults
• Review findings (numbers, docs claims, paper) addressed; Figure 1 panels now use line styles as well as color