jacobian='jax' emits jnp.jnp.log for any model using ln(): double substitution in _translate_expr_jax

Read python/bngsim/_jax_rhs.py around lines 143-144 and run the minimal _translate_expr_jax reproduction. Check python/tests/test_jax_jacobian.py for the expression-translation tests, then add coverage for ln() so it produces jnp.log once and verify the JAX run matches the default trajectory.

Summary

_translate_expr_jax applies re.sub(r"\bln\b", "jnp.log", c) on line 143 and then re.sub(r"\blog\b", "jnp.log", c) on line 144. The second substitution re-matches the log it just produced (the . in jnp.log is a non-word char, so \blog\b has a word boundary there), so ln(x) becomes jnp.jnp.log(x). The generated source is eval'd in rhs() with {"__builtins__": {}}, so the very first RHS evaluation raises AttributeError. Any .net model with an ln() in a functions block is broken on the JAX paths: bngsim.Simulator(model, method='ode', jacobian='jax', net_path=...) and bngsim._diffrax_solver.run_diffrax, which both call generate_jax_rhs. The same model integrates fine with the default jacobian, so this is purely the translator's substitution ordering.

Location: python/bngsim/_jax_rhs.py:143

Reproduction

cd /private/tmp/claude-11919/-Users-hlavacek-Code-bngsim/c49bfe51-f320-4eef-88bb-4025a472c19f/scratchpad && /Users/hlavacek/Code/bngsim/.venv/bin/python verif_ln_e2e.py    # (model at verif_ln_model.net, identical content to the claim's int_ln_model.net). Minimal one-liner that needs no .net file: /Users/hlavacek/Code/bngsim/.venv/bin/python -c "from bngsim._jax_rhs import _translate_expr_jax; print(_translate_expr_jax('k1*ln(1+B)', {'k1':0}, {'B':0}, set(), []))"  ->  params[0]*jnp.jnp.log(1+obs[0])

# int_ln_model.net (written to the scratchpad dir):
# Created by BioNetGen 2.9.3
begin parameters
    1 k1     0.5  # Constant
    2 kd     0.1  # Constant
end parameters
begin functions
    1 growth() k1*ln(1+B)
end functions
begin species
    1 A() 10
    2 B() 1
end species
begin reactions
    1 1 2 growth #_R1
    2 2 1 kd #_R2
end reactions
begin groups
    1 A_tot                1
    2 B                    2
end groups

# int_ln_e2e.py
import warnings, bngsim
NET = "/private/tmp/claude-11919/-Users-hlavacek-Code-bngsim/c49bfe51-f320-4eef-88bb-4025a472c19f/scratchpad/int_ln_model.net"
m = bngsim.Model.from_net(NET)
with warnings.catch_warnings():
    warnings.simplefilter("ignore")
    sim = bngsim.Simulator(m, method="ode", jacobian="jax", net_path=NET)
try:
    r = sim.run(t_span=(0, 10), n_points=11)
    print("ran ok", r.species[-1])
except Exception as e:
    print("E2E RAISED:", type(e).__name__, e)

Observed

$ .venv/bin/python int_ln_e2e.py
E2E RAISED: AttributeError module 'jax.numpy' has no attribute 'jnp'

$ .venv/bin/python -c "from bngsim._jax_rhs import _translate_expr_jax; print(_translate_expr_jax('k1*ln(1+Btot)', {'k1':0}, {'Btot':0}, set(), []))"
params[0]*jnp.jnp.log(1+obs[0])

$ .venv/bin/python -c "..."   # same model, default jacobian
baseline (jacobian=auto) OK: [ 0.84228104 10.15771896]

Expected

ln(x) should translate to jnp.log(x) and the JAX-Jacobian run should produce the same trajectory as the default jacobian ([0.84228104, 10.15771896] at t=10). Fix by making the substitutions non-re-entrant, e.g. map ln and log in a single pass (re.sub(r"\b(?:ln|log)\b", "jnp.log", c)) or use a placeholder/\b(?<!\.) guard so an already-emitted jnp.log is not rewritten.

Verification notes

The repro was independently re-run and the analysis re-checked against existing tests, git blame/git log -S and CHANGELOG.md (confidence: high).

The claim survives every refutation I could mount.

MECHANISM CONFIRMED BY READING. /Users/hlavacek/Code/bngsim/python/bngsim/_jax_rhs.py:143-144:
c = re.sub(r"\bln\b", "jnp.log", c)
c = re.sub(r"\blog\b", "jnp.log", c)
The second sub re-matches the log the first one just emitted (. is a non-word char, so \blog\b has boundaries inside jnp.log), giving jnp.jnp.log. No other entry in that block is re-entrant: exp->jnp.exp, min->jnp.minimum, etc. never re-match, which my probe confirms — only ln is affected.

NOT INTENDED. The function's own docstring at _jax_rhs.py:95 states the contract it violates: "- ln() -> jnp.log()". There is no test asserting the doubled form: python/tests/test_jax_jacobian.py's TestExpressionTranslation covers only simple/time/exponentiation cases (lines 219-246) and no test file mentions ln at all. git blame -L 141,146 puts both lines in the file's original commit 764f4b1; git log -S finds no later commit that deliberately shaped this. CHANGELOG.md has no entry for it (its _jax_rhs.py mentions are about dense-solver routing and a de-duplicated expression, not the math-function map).

INPUT IS VALID, NOT MISUSE. ln is a first-class .net function everywhere else in this codebase: src/bngsim_api.cpp:191 and src/expression.cpp:1191 list it, src/expression.cpp:961 registers ln in the symbol table, and python/bngsim/_jacobian.py:121, _saturable_jacobian.py:58-59, _codegen.py:5921 all map ln->log. So the whole rest of the system consumes ln( correctly — CHANGELOG.md:3871 even quotes a real run's logterm() = ln(1 - Atot). My identical .net integrates cleanly on the default jacobian ([0.84228104, 10.15771896]); only the JAX translator breaks on it.

NOT A DELIBERATE FALLBACK / NOT SWALLOWED. The jax branch in python/bngsim/_simulator.py:1060-1087 raises on every other failure mode (no net_path -> ValueError, no jax -> ImportError) and has no try/except around prepare_jax_jacobian, so the AttributeError propagates out of the public run. I also confirmed the second, independent entry point fails: generate_jax_rhs(net_path) — used by python/bngsim/_diffrax_solver.py:95 — raises the same AttributeError.
… (truncated)

Severity assessed as medium. Reviewed against main at commit c3ab2ee.