theory_fn protocol, default CCL backend, pure order-independent shift draw

[cailmdaley]

End state

Blinding is driven by a single theory-function protocol instead of a firecrown likelihood. ConcealDataVector obtains both theory vectors (fiducial and hidden) by calling a theory_fn(cosmo_params) -> np.ndarray; the fork ships a default CCL cosmic-shear backend so a standard SACC blinds out of the box, and a caller can supply any callable. import smokescreen imports no theory backend — pyccl enters sys.modules only when the default backend is constructed, and firecrown is never imported at module scope anywhere in the package.

Closes #2

Interfaces

ConcealDataVector(fiducial_params, shifts_dict, sacc_data, *, seed, theory_fn=None, shift_distr="flat")

• fiducial_params is a plain mapping of cosmological-parameter name → value (CCL-native names); it replaces the old cosmo + likelihood pair. The hidden point is fiducial_params with the drawn deltas overlaid — no pyccl.Cosmology object, no _create_concealed_cosmo.
• theory_fn=None builds the default CCL backend from sacc_data; supplying a callable overrides it. No systm_dict — systematics, if a backend needs them, are closed over inside its theory_fn.
• seed is keyword-only with no default (blinding custody depends on a deliberate secret seed). int or str; a string is normalized to an int via SHA-256.
• The sole construction guard is a length check (len(theory_fn(fiducial_params)) == len(sacc_data.mean)); the value/covariance _verify_sacc_consistency check is dropped, not re-expressed.
• calculate_concealing_factor("add"|"mult") computes theory_fn(concealed) - theory_fn(fiducial) (or the ratio); apply_concealing_to_likelihood_datavec() returns sacc_data.mean + factor (or * factor).
• save_concealed_datavector deep-copies the SACC, overwrites its mean, carries the covariance over unchanged, stamps custody metadata + seed, and writes through SACC's own save_fits/save_hdf5 — no firecrown, no save_to_sacc/get_sacc_indices.

draw_param_shifts(shifts_dict, seed, shift_distr="flat") -> dict[str, float] (smokescreen.param_shifts)

• Pure, order-independent (sorted keys, one scalar draw per key via a local numpy.random.default_rng; never touches process-global np.random).
• Returns deltas to add to the fiducial, not absolute values. float h → symmetric envelope (-h, +h); (lo, hi) tuple → delta box. "gaussian" accepts only the float form.
• Seed int used as-is; str reduced to a 64-bit int via SHA-256 of its UTF-8 bytes (stable across runs/machines).

Default CCL backend (smokescreen.backends.ccl.build_ccl_theory_fn)

• The only import pyccl in the package; imported when the backend is constructed. Given a SACC with weak-lensing n(z) tracers and cosmic-shear rows (galaxy_shear_cl_ee and/or galaxy_shear_xi_plus/xi_minus), it builds a WeakLensingTracer per bin and returns the Cℓ / ξ± vector aligned to sacc_data.mean. Two documented conventions of this convenience backend (its row-to-theory contract): it defaults to CCL-native eisenstein_hu + halofit so it runs on a bare pyccl install with no Boltzmann code (overridable via cosmo_params), and it reads the ξ± theta tag as arcminutes (converted to degrees for pyccl.correlation).

Inherited firecrown path (smokescreen.firecrown_datavector.FirecrownConcealDataVector)

• Upstream's firecrown-likelihood concealment class, retained so the fork doesn't amputate the upstream integration. All firecrown/pyccl imports are function-local (lazy); the module docstring marks it inherited-from-upstream and unsupported in this fork. Nothing in the package's main modules imports it, no test exercises it, and firecrown is not declared or installed here. The firecrown cosmic-shear example likelihood under examples/ is likewise untouched.

CLI (smokescreen.__main__)

• datavector reworked onto the new surface: datavector_main(path_to_sacc, fiducial_params, shifts_dict, seed, shift_type='add', shift_distribution='flat', path_to_output=None, keep_original_sacc=False, output_suffix=None), driving the default CCL backend (the firecrown-specific likelihood_path / systematics / reference_cosmology params are gone). encrypt / decrypt are unchanged.

Test coverage

The suite runs firecrown-free (pyccl present) — 57 passed, nothing skipped for a missing firecrown.

• Blinding mechanics against a synthetic theory_fn closure (importing neither pyccl nor firecrown): construction, concealing factor equals the theory difference bit-for-bit for "add" and the ratio for "mult", apply returns mean ± factor, save writes the blinded mean with the covariance untouched and the custody metadata stamped, and the length guard raises on a mismatch.
• Shift draw directly: fixed (shifts_dict, seed) reproduces a fixed shift; permuting the key order yields identical per-key deltas; the draw leaves np.random's global state untouched; the drawn value is a delta within the envelope (not an absolute); a string seed and its normalized int agree.
• Import discipline via subprocess: after a fresh import smokescreen (and import smokescreen.datavector), neither pyccl nor firecrown is in sys.modules; pyccl appears only once the CCL backend is built.
• Default CCL backend end-to-end on a small in-test cosmic-shear SACC (2 n(z) tracers, ξ±): blinds to a vector that differs from the input with the covariance untouched.

Notes for review

• _verify_sacc_consistency is intentionally absent (row-to-theory alignment is the backend's contract; with no likelihood there is no second internal vector to compare against). This is a deliberate, non-behavior-preserving change.
• Packaging, pyccl dependency declaration, CI wiring, and install docs are out of scope here (separate packaging sub-PR); pyproject.toml, CI workflows, and README are untouched.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KpaRHk3QwN13myduQ3hJyf