RL: unify all randomness under a single random_seed (seed audit)

[rutayan-nv]

Summary

cloudai should have one seed that governs all randomness, derived from the canonical BaseAgentConfig.random_seed (default 42). Today several RNG sites are seeded independently — or not at all — so a run is not reproducible end-to-end from a single knob. This issue records an audit of every randomness source in cloudai core and proposes threading them all from the single seed. Audit only — no fix is requested in this issue.

Motivating bug (already addressed in the env_params/DR work): the domain-randomization sampler defaulted to seed 0 while the agent default is 42, so a config without an explicit random_seed ran the agent and the env on two different seeds.

Canonical seed

• BaseAgentConfig.random_seed: int = 42 — src/cloudai/configurator/base_agent.py. Single declared knob, set via TOML [agent_config] random_seed; every agent config inherits it.

Audit of RNG sites in cloudai core

#	Site	RNG	Seeded from random_seed?	Notes
1	DR sampler — env_params.py (random.Random(f"{seed}:{name}:{trial}"))	Python random	Yes (after the DR fix)	Per-(seed, param, trial) sub-seeding keeps params independent and reproducible.
2	Gym reset(seed) / seed() → np.random.seed(seed) — cloudai_gym.py	numpy global	No (only when a caller passes seed)	In the native BaseAgent.run() loop, env.reset() is called with no seed, so numpy's global RNG is never seeded from random_seed.
3	start_action: "random" — base_agent.py	n/a in core	No	The contract implies a random first action, but core provides no RNG for it; GridSearchAgent ignores it.
4	GridSearchAgent — grid_search.py	none (itertools.product)	n/a	Deterministic; the only agent shipped in core.

Gaps

• G1 — no single propagation path. Each RNG (Python random, numpy global, gym reset, plus external-agent frameworks) is seeded independently; nothing guarantees they all derive from random_seed.
• G2 — numpy global RNG unseeded by default. The gym only seeds numpy when a gymnasium caller passes seed; the native cloudai loop never does, so in-process np.random usage is nondeterministic across runs.
• G3 — start_action="random" is unwired in core. No core RNG tied to random_seed selects the random start.

Related (out of core)

Stochastic agents (BO, GA, MAB, PPO, DQN) live in cloudaix, not core. Each must read agent_config.random_seed and seed its own framework (numpy / torch / RLlib). There is no central enforcement that they all use the same single seed; this should become a documented contract once core exposes a single derivation utility.

Proposed direction (non-binding)

• Add a single seed-derivation helper in core, e.g. derive_seed(random_seed, component: str) reusing the f"{seed}:{component}" hashing the DR sampler already uses.
• Thread it to: the DR sampler (done), the gym (seed numpy global from random_seed when no explicit seed is given), and start_action="random".
• Publish a contract for external agents: seed your framework from derive_seed(random_seed, "<agent>").

Acceptance criteria

• One knob (random_seed) makes a full cloudai run reproducible — DR draws, any numpy usage, and random start actions.
• No RNG site defaults to a seed that disagrees with random_seed.
• Independence preserved via per-component sub-seeds (not one shared global stream).