[wshlavacek]

Closes #996. Closes #997. Closes #999.

Each of these configurations ran to a result with no meaning. Each is now refused before anything runs, with a message that names the cause.

#996: normalization = scale under a Bayesian sampler

• scale profiles a multiplicative scale out at each evaluation. That is a profile, not a marginal, so the draws were not posterior draws. noise_profiling and linear_profiling were already refused for samplers for this reason.
• A new Configuration._bayesian_sampler(), the registry family 'sampler' (am, dream, hmc, mh, p_dream, pt), now decides this for all three refusals.
• Covered forms:
  • edition 2: whole-fit, per-observable, per-experiment and in a chain;
  • edition 1: whole-fit.
• The message points to a measurement-layer scale (observable: …, formula: c * y with c free) or an optimizer.
• Oracle: PyBNF's profiled objective against a quadrature over c. The gap equals ½ log(2π/A), which moves with the prediction.

#997: normalization on a column scored by a count family

• Normalized values are not counts. The PMF's continuous extension is not a likelihood, and the fit moved with the normalizing constant.
• A new NoiseModel.scores_counts property (True for NegBinomial) decides this, not the class name, and it covers init, peak, zero, unit and floor.
• scale keeps fix(am): record every experiment's trajectory, normalized once; refuse what would be wrong (#990) #994's "no additive scale" refusal, so each case gets one message.
• The message names the experiment and condition.
• A normalized Gaussian column next to a raw neg_bin column still loads.
• Oracle (scipy):
  • at integers the score equals -nbinom.logpmf and sums to 1;
  • the continuous extension integrates to 0.68 and 0.82;
  • a fitted θ moves from 7.10 to 7.75 to 7.92 with the normalizing constant.
• Non-integer raw count data is left open. Across the 361 example configs that use neg_bin, no scored column holds a non-integer, so nothing in the repository depends on it.

#999: network-free experiments that BNG2.pl cannot run

• What BNG2.pl cannot run (grid.bng2pl_network_free_refusals, checked against BNG2.pl 2.9.3 itself):
  • method nf without n_steps, which covers the edition-2 time course (sample_times) and dose-response;
  • any network-free method other than nf, i.e. RuleMonkey. That is the BNG2.pl half of Nine loud failures that block a valid job or give the wrong cause: RuleMonkey experiments, one-time-point data, PEtab import and export cases, error messages that hide their reason, and a crash on deeply nested formulas #962 item 1.
• When it is refused:
  • at load under bngl_backend = bionetgen and under job_type = check, which always runs BNG2.pl;
  • at set-up under auto when the model would land on BNG2.pl: bngsim unavailable, bridge can't take it, missing engine, or a failed set-up that fell back.
• auto is not refused at load. Whether it uses bngsim depends on the machine, and readers such as PEtab export simulate nothing.

Sibling paths

Path	#996	#997	#999
edition 1	refused (whole-fit)	refused	same check, hand-written and param_scan actions
edition 2	refused (all forms)	refused	refused
PEtab import	cannot arise	cannot arise	goes through the same refusal (tested)
-r resume	refused (tested)	refused (tested)	refused (tested)
job_type = check	not a sampler	refused	refused on any backend

Tests

• New file: tests/test_load_refusals_996_997_999.py, 55 tests.
• Existing tests changed:
  • test_network_free_bngsim_supported::test_nf_under_bng2pl_builds asserted the combination An edition-2 network-free experiment (method: nf) under bngl_backend = bionetgen fails every simulation, because BNG2.pl's NFsim refuses the sample_times PyBNF writes, and nothing refuses it at load #999 refuses, and its comment wrongly said BNG2.pl's NFsim honours the output points. It now asserts the refusal.
  • Five fix(am): record every experiment's trajectory, normalized once; refuse what would be wrong (#990) #994 tests built neg_bin + normalization or am + scale. They now assert the load-time refusal, and am's own checks are exercised on Gaussian columns.
• Example-config sweep: none of the configs that loaded before is newly refused. 364 configs fail to load for reasons that predate this PR: comma paths, cluster bng_command, and space-separated data files. Those will be filed separately.
• Runs: about 25 related test files pass, each run on its own, on the bngsim main build. ruff and tools/changelog.py check are clean.

Review

The first independent review found no blocking defect and checked that nothing valid is refused:

• legacy network-free runs on BNG2.pl, and check;
• per-observable noise mixes;
• profiling refusals, unchanged across all 22 job types;
• the example sweep (549/553 load, identical to main).

It raised four findings, fixed in 47c21b6:

• F1: the predicate missed hand-written lines BNG2.pl stops (trailing ; or comment, simulate_nf, method=>"NF", n_steps=>0, continue=>1, no t_end). Lines are now read as BNG2.pl runs them.
• F2: the scale hint suggested one shared c. It now asks for one per profiled series.
• F3 and F4: wording.

The reviewer's tests are committed (tests/test_load_refusals_review_1001.py). A second check of 47c21b6 tried 24 hand-written network-free actions against BNG2.pl itself, and none that BNG2.pl runs is refused.

b52d08c isolates one test from the bngsim-less CI leg's PYBNF_NO_BNGSIM.