ENH: seed Monte Carlo per simulation index for worker-invariant inputs

[thc1006]

This makes Monte Carlo input sampling reproducible regardless of how the run is executed. Today the stochastic models are seeded once per worker (parallel) or once before the loop (serial), so the inputs you get depend on the worker count and on which worker draws which simulation. Two runs with the same seed but different worker counts produce different inputs.

The change seeds per simulation index instead. From the run's root seed I spawn one SeedSequence child per index, and reseed the models from child_seeds[i] before simulation i. SeedSequence.spawn is invariant to the spawn count, so index i always maps to the same child, including in append mode. Each index seed is then split three ways so the environment, rocket, and flight don't share a draw stream.

After this, the generated inputs are identical across serial, parallel(2), and parallel(N) for a fixed seed, and a split append run (3, then +3) matches a single run of 6.

Behavior changes worth calling out

Both shift the numbers a fixed seed produces, so any stored baseline regenerates:

1. The env/rocket/flight decorrelation. They used to share one seed, so their first draws were correlated.
2. The serial index field now counts from 0, matching the parallel path, which already did.

Tests

New tests/unit/simulation/test_montecarlo_determinism.py: serial reproducibility, worker-invariance (serial == parallel(2) == parallel(4), bit-identical per index), append reproducibility, and the None-seed path. They stub Flight so only input sampling runs, and drop the one stdlib-random thrust_source so the inputs are byte-reproducible from the numpy seed alone. Marked slow (each sim rebuilds a rocket), matching the existing test_monte_carlo_simulate.

Locally: ruff and pylint both 10/10, and the four tests pass with --runslow.

Known gap

List-valued stochastic attributes are drawn with the stdlib random.choice (a global, unseeded generator), so a model with multi-element lists isn't fully worker-invariant yet. BPC isn't affected (its lists are single-element). A general fix would seed the stdlib random per index too, or move those draws onto numpy. I'll note it upstream when this goes to RocketPy.

Unrelated, but flagging it since I ran the suite: test_monte_carlo_simulate is already failing on develop (processed_results has no "apogee", and the sims themselves don't error), independent of this change. Happy to look at that separately if it's useful.

[zuorenchen]

Let's pull this from RocketPy once RocketPy-Team#1054 is merged

[thc1006]

Let's pull this from RocketPy once RocketPy-Team#1054 is merged

for sure
but upstream not yet merged should we wait or implement it ahead of them

[zuorenchen]

I'd prefer to wait for them to reduce conflicts when pulling changes from RocketPy in the future. Also, we'll probably stay with serial monte-carlo for now

[thc1006]

Marking this draft to match what you asked for on this PR: wait for RocketPy#1054 and pull the implementation back through the usual sync. It is still open, at 8 commits and 9 files, so there is nothing to pull yet.

Leaving it open and unmarked was costing review attention it should not be: this branch is one commit from July 18 and is now the older, narrower version of work that moved upstream. Two things worth recording before it sits, since they are true of this branch and would matter if anyone picked it up later:

• The index claim here is not atomic. keep_simulating() and increment() are separate manager calls, so two workers can both pass the final check and one can index past the end of the seed list. ENH: reproducible Monte Carlo via per-simulation-index seeding RocketPy-Team/RocketPy#1054 holds the lock across both.
• Passing SeedSequence objects into _set_stochastic() reaches CustomSampler.reset_seed(), whose documented input is an integer, so an integer-only sampler such as random.Random breaks. ENH: reproducible Monte Carlo via per-simulation-index seeding RocketPy-Team/RocketPy#1054 converts each stream to an integer.

Neither reaches BPC today, since it runs Monte Carlo with parallel=False.

One thing that does not depend on this PR and is worth its own issue at some point: StochasticModel.dict_generator and StochasticRocket._randomize_position sample list-valued attributes through the process-global random.choice, which no seed controls. That is on develop now, not something this branch introduces, and it means a model with a multi-element list is not reproducible from a seed even in serial. I have not measured whether BPC's balloon model has one; I will check and open an issue if it does.