fix: event-time sensitivity guard misses state-dependent triggers reached through SBML (#52)

[wshlavacek]

Closes #52.

The bug

The guard refuses forward sensitivities when an event's crossing time depends on a requested sensitivity parameter, and it decides that from the trigger's bound addresses. It compared them against species concentrations only.

ModelBuilder registers a species as an ExprTk variable only when its name is still free:

// src/model_builder.cpp — species names may collide with observable names
try { eval.define_variable(sp.name, &sp.concentration); }
catch (...) { /* already registered (e.g. as observable) — skip */ }

SBML models routinely give each species an observable of the same name, so that registration is skipped and a trigger token binds to &obs.total, never to &sp.concentration. Every SBML state-dependent trigger therefore slipped the guard and was answered instead of refused, with the event contributions missing entirely.

Confirmed directly on the reported shape — the SBML_EVENT fixture with its trigger swapped to S < 5:

species names: ['S']
obs names:     ['S']          <- the collision
guard reason:  None           <- no state dependence seen
!!! ANSWERED (should have been REFUSED)

On AMICI's neuron fixture (Izhikevich, trigger v > 30, which names no parameter but whose crossing time depends on a and b through the trajectory) the issue measured sensitivities 6x–135x off, uniformly in one direction, across all four parameters — while dirac and nested_events, whose triggers name a parameter outright, were correctly refused.

The fix

The guard now tests against every address that carries live state: species concentrations, observable totals, and rateOf accessors. An observable total is a linear functional of the state and a rateOf accessor is dx/dt, so a trigger reading either has a non-zero dt*/dp exactly as a concentration read does.

Refusing is unchanged as a policy — as the issue says, "the codebase already does it well" — this is only the coverage of what counts as state. The message now names which of the three it saw and why that implies a moving crossing time.

Tests

• An SBML model whose trigger reads a species (the neuron shape) must be refused, through both run and compute_all_sensitivities. Both fail on the unmodified core with "DID NOT RAISE".
• The test asserts the species/observable name collision explicitly, so if that binding ever changes the test says why it was written rather than silently becoming vacuous.
• No false positives: the existing ModelBuilder-path state-dependent test still passes, as do fixed-time events, the parameter-valued reset, compute_all_sensitivities on a fixed-time model, and the discontinuity-trigger pulse model (n_events == 0, must stay allowed).

Full Python suite: 2575 passed, 89 skipped.

Note on the issue's item 3

The issue separately asks whether the jump terms are handled at a fixed event time — neuron assigns v = -c and u = d + u, contributing d/dc and d/dd independent of dt*/dp. That is orthogonal to this guard and already covered: test_parameter_valued_reset_matches_fd pins ∂h/∂p against finite differences for a parameter-valued assignment, and test_additive_bolus_is_continuous pins the ∂h/∂x = 1 case.