chaos

2. Key Technical Concepts:
   • Chaos vs ill-conditioning discriminators: chaos = divergence grows unbounded/$\epsilon$-independent (sensitive dependence); ill-conditioning = bounded near-singular amplification. THE decisive test = decay ratio D_final/D_max (chaos cannot decay). Amplitude law $D_{\mathrm{sat}}\propto \epsilon \Rightarrow$ B; $\epsilon$-independent $\Rightarrow$ chaos (but degenerate when perturbation washes out — reads noise floor).
   • The verdict: NOT chaos. Bounded near-singular ill-conditioning (B). Perturbations WASH OUT (same mechanism as the startup-posture q0 seed). The §4.6 Lyapunov certificate ($\dot{V}\le 0$) makes the operating region provably non-chaotic.
   • dt derived-param-at-load bug: parameter_loader bakes dt-dependent params (EE-goal smoothing alphas via exp_scale(dt, time_scale), startup.steps_effective = ceil(startup.time/dt), com smoothing.alpha) during finalize_parameters() BEFORE pre_run_loader.apply_overrides runs. Same class as the gain-built-matrix footgun. Fix: re-derive on a dt override (mirrors the gain-rebuild special-case).
   • The dispatcher: dispatch.py run --manifest <file> syncs cluster→controller HEAD (credential-free thin bundle), preflights (drops dirty/stale boxes), partitions stems by weight, launches detached tmux, polls, collects with provenance. A SWEEP stem (chaos = 4 eps) runs all variants in ONE Orchestrator process on ONE box (within-box → no cross-box divergence confound). Collection nests results under run/ + stem/ subdirs.
   • Coverage marking: passive in POSE mode (the adopted default — scorer inert); active in ANCHOR (feeds target selection). dt_refinement runs POSE → marking is passive → doesn’t affect p_c/p_e/s_min_J.
   • pytest discovery: tests auto-discovered by name (test_*.py files, test_* functions). “Suite” is emergent = all test_*.py under validation/. pyproject.toml [tool.pytest.ini_options] with testpaths + markers. @pytest.mark.slow + -m "not slow" for a fast subset. pytestmark is a SINGLE module variable — multiple assignments clobber; use a LIST.