reseed_chain() on a greedy chain removes the greedy sampler and adds dist — deterministic chain silently becomes stochastic

[lloyal-research]

sampler::create_chain() ends the chain differently depending on temperature:

• temp > 0 → … → temp → dist(seed)
• temp ≤ 0 → … → greedy (no dist)

sampler::reseed_chain() assumes the first shape:

// Remove last sampler (dist by convention)
llama_sampler* old_dist = llama_sampler_chain_remove(chain, n - 1);
...
llama_sampler_chain_add(chain, llama_sampler_init_dist(new_seed));

Called on a greedy chain, this removes the greedy sampler and appends dist — converting a deliberately deterministic chain into a stochastic one, silently. No error, no log; the chain shape simply changes.

BranchStore tracks sampler_has_dist (set from temperature > 0 at creation) precisely so its callers can avoid this — but the guard lives in the caller's flag, not in the function that has the hazard. Any direct user of reseed_chain, or any future call site that forgets the flag, corrupts the chain.

Smallest fix

Make reseed_chain self-guarding: check whether the last sampler is dist (or accept a has_dist parameter) and no-op — or log loudly — on a greedy chain. The flag on BranchStore can stay as an optimisation; correctness shouldn't depend on it.

Found during a full read of sampler.hpp/branch.hpp; one of five silent-failure edges filed together.