[Montana]

Closes #1400: both items, the wording and the carried column.

What was wrong

bngl_capability_report refused every model with a species an SBML assignment rule sets, saying "the flat .net round-trip cannot carry the rule ... (GH #205/#223)".

• The rule was carried. The writer emits it as the function of the species' own name (B() 2.0*__sp1), and every rate law reads that, so the other species' dynamics are exact.
• Only the species' own column was lost. It is written $-fixed, and BNG2.pl holds it at its initial value, as does bngsim reloading BNG2.pl's .net. That .net names the species as a pattern (@comp1::$B()), which never equals its rule's name, so _ar_report_map_from_net (sbml_to_net's default L2 gate certifies a species assignment rule that the reloaded .net never applies; validate="full" catches it #515) never found the pair.
• The citations pointed elsewhere: The codegen .so cache grows without bound and has no supported way to inspect, prune, or clear it (2.0 GB / 14,377 entries on one machine) #205 is the codegen cache, test(audit): report corpus-absence skips as one number with a remediation (#222) #223 a skip-audit PR.
• The bng2 gate could not see it: its probe is the right-hand side, and a held species has dy/dt = 0 on both sides.

The fix

Item 1, the wording. The note and its comment state the real mechanism, the stale citations are gone, and the test module's docstring and refusal row are corrected.

Item 2, the column is carried, as the issue sketches:

• Writer. A rule species' molecule is named __ar_<rule> (AR_MOLECULE_PREFIX), the name of the function or observable that carries its rule: @comp1:$__ar_B() beside B(). _molecule_names takes the carriers, so the writer and the round-trip's aligner (_name_aligned_perm) build the same names. A species whose mark would collide with an existing name is left unmarked and keeps the refusal.
• Reader. _ar_report_map_from_net pairs a fixed @comp::__ar_X() with the function or non-identity observable X. An unmarked fixed species keeps the exact-name rule it had: a BNGL species name is a pattern, and only the mark says it is a rule's. An unfixed marked species is not paired.
• Gate. sbml_to_bngl(validate="bng2") now also runs _ar_report_delta, which gains back_names and back_perm because BNG2.pl's names and order are not the source's. The report carries max_ar_report_delta, and rhs_faithful requires both deltas, as on the .net path. roundtrip_deltas returns all three; roundtrip_rhs_delta keeps its signature.
• Policy. A marked species is a bngsim_only note, warned and never refused, as net writer: warn when a rule species is written so only bngsim reads it as meant (#1274) #1345 made it on the .net path, because BNG2.pl's own run still holds the column. The docs say so (interchange.md).
• New refusal. A model with no reactions is now refused by name. BNG2.pl's generate_network aborts on one ("Nothing to do: no reaction rules defined"). BIOMD0000000566 and 567 are rule species alone: refused for those before, they would now be written as files that do not build.

Measured

The issue's models, write_bngl → BNG2.pl generate_network → Model.from_net, against Model.from_sbml (t = 0-5, rtol 1e-10):

model	rule column, main	this branch
B := 2*A, C reads B	held at 20 (source 1.64)	1.64169998, every column to 1e-9
B := A*A	held at 100	exact
$B consumed by J3: B -> D; k3*B	held	exact
B := 2*A in a compartment of volume 2	held	exact

The corpus: every model under rr_parity/models with a rule species (289), through sbml_to_bngl(validate="bng2", strict=False):

• 175 are no longer refused for their rule species.
  • 174 pass the gate: the rule column within 6.8e-14 and the right-hand side within 6.8e-11.
  • 1 is refused by the gate: BIOMD0000000664 reports g2_0 as 1 at t = 0 in the source and 0 in the reload, and both agree afterwards. Under strict it is refused, as it should be.
• 80 are still refused for other reasons, none of them for a species that could not be marked.
• 34 do not load or build, as on main: 26 under-specified, 6 model errors, and the two reaction-less models above, which this PR now refuses up front.

Tests

test_bngl_rule_species.py (11), every one failing on main:

• the species is written as __ar_B beside B() and warned, not refused;
• a species that cannot be marked is still refused;
• a model with no reactions is refused;
• the reader pairs only a marked, fixed species: not an unmarked $Y() beside Y(), nor an unfixed __ar_Z();
• the column check reads a reordered reload in its own order;
• through BNG2.pl: every column, the rule's included, comes back on all four of the issue's models; the gate checks the column and passes; and BNG2.pl's own simulate still holds it (.cdat: 20.0, the others to 1e-6). That last one flags the bngsim_only note as stale if BNG2.pl ever carries the column.

test_sbml_to_bngl.py:

• test_validate_bng2_gate patches the new seam and adds a forced rule-column divergence, which raises as an RHS one does;
• BIOMD0000000009 leaves test_refuses_fail_loud, since it is no longer refused for this.

Mutations

mutant	fails
N1: the reader ignores the mark	6
N2: the writer does not mark	7
N3: the aligner without the marks	the gate test
N4: the gate ignores the column	test_validate_bng2_gate
N5: the column check ignores back_perm	the reorder test
N6: the mark ignores collisions	the cannot-mark test
N7: the reader pairs an unfixed marked species	the reader test

Checked

• Full suite: 16,849 passed, 138 skipped. pre-commit run --all-files and sphinx -W are clean.

Not in this PR

• BNG2.pl's own run still holds the column (its .gdat carries the rule's value with print_functions=>1). Only bngsim reads the pair back; that is the bngsim_only note.
• BIOMD0000000664's t = 0 report of an alias rule (g2_0 := g2): the gate refuses it, and why the source reports 1 there is a question of its own.