[wshlavacek]

#1383's search for smooth pulses in time (#1286) made bngsim slow on any model whose rate laws call exp, cosh or tanh, on every run. Most such calls have nothing to do with time: a Boltzmann factor, a rate of a voltage. The search inlined every rate law and parsed each one that holds such a call, before asking whether any of them involves a clock.

model	search per run, before	after
MODEL1006230090	over 7 minutes (killed)	8 ms
MODEL1006230049	21 s	under 20 ms

It found no pulse in any of them. On the nightly's runners that pushed MODEL1006230049, 053 and 077 out of their budget in the path-agreement check. #1401's run 41313484937 timed all three out in the locale check, which main before #1383 passed in 1,350–1,456 s. The slowdown never shipped: #1286's changelog fragment is still unreleased, so this adds none.

The fix

A pulse's argument has the clock as its only symbol once the parameter values are put in. The search now reads the function texts first. For each call's argument, it follows the functions and derived parameters it names and asks two things: does the argument reach a clock (time(), or a counter fed at a constant rate), and does it reach any other species or observable? Only an argument that reaches a clock and nothing else can be a pulse. If no argument passes, nothing is inlined or parsed. Otherwise the search runs exactly as before.

Checks

• Tests. Four new cases fail without the change: a Boltzmann factor, a pulse of the state, the clock reached only beside the state (as pH_cy does in MODEL1006230049), and _e^ of no clock. Two more confirm that a pulse reached through another function is still found, in time() and in a counter. ode: stop on the centre of a smooth bump of the clock in a rate law (#1286) #1383's existing 30 tests pass unchanged.

• Every pulse ode: stop on the centre of a smooth bump of the clock in a rate law (#1286) #1383's tests find is still found. I ran those tests and test_net_state_switch_stepping.py with a guard that asserted the check never skips a search that finds a stop. 75 searches ran: all 44 that found a stop were kept, and the 27 skipped found none.

• Corpus sweep, all 1,319 BioModels SBML files. For every model the check skips, the old full search was run too.

  	models
  no exp/cosh/tanh call at all (untouched)	1,109
  searched as before	86
  skipped, and the full search finds no pulse	91
  skipped, full search past 300 s (MODEL1006230090)	1
  load error (unrelated)	32

  None skipped had a pulse. MODEL1006230090 has no clock at all: no function calls time(), and no counter. The check costs at most 8 ms per model.