fix(exp12): deterministic per-(seed,trial,bit) sampler and fixture path to restore sim tests

[jverdicc]

Motivation

• The exp12 simulation used a single shared sequential RNG stream across scenarios which made results scenario-order dependent and broke the monotonicity property test under topic budgeting.
• The golden-fixture lookup in the test used a repository-relative path that could fail depending on the test working directory.

Description

• Replace the shared LCG stream with a deterministic per-(seed, trial_idx, bit_idx) sampler implemented as uniform_u53(seed, trial_idx, bit_idx) and update binomial_sample to accept seed and trial_idx so scenario ordering is independent.
• Remove the shared Lcg64 state and call binomial_sample(..., cfg.seed, trial_idx) for each trial.
• Make the test robust by resolving the fixture via env!("CARGO_MANIFEST_DIR") when reading test_vectors/exp12_golden.json.
• Update the golden fixture crates/discos-core/test_vectors/exp12_golden.json values to match the corrected deterministic output.

Testing

• Ran cargo test -p discos-core --features sim --test exp12_tests and the test binary passed: both exp12_matches_golden_fixture and exp12_properties_hold succeeded.
• Ran cargo test -p discos-core --features sim and the package tests completed successfully.
• Attempted workspace checks: cargo fmt --check, cargo clippy --workspace --all-targets -- -D warnings, and cargo test --workspace were executed but surfaced pre-existing unrelated issues in other crates (formatting, lints, and CLI API/compile drift) and are therefore not blocking this exp12 fix.

---

Codex Task