Per-element parameterization: mix and match constraints per instance (8.2.1)

[jdeast]

Mix and match parameterizations per instance: Parameter.build_pymc no longer derives a whole vector or none of it. Review item 8.2.1.

Why

Four shipped features wanted a per-instance choice and could not have one, because is_derived = np.full(n_elements, expr_raw is not None):

• band.ld_law raised on a system mixing quadratic and linear bands (workaround: quadratic everywhere with q2 pinned at 0.5, at the cost of a prior uniform in q1 rather than u1);
• planet.mass_parameterization raised on an explicit per-planet disagreement and silently fell back to all-linear on an implicit one -- handing an unconstrained planet a uniform prior over [-1000, 260000] Mjup instead of the ratio coordinate its topology asks for;
• star.mist/parsec declared a mask on age/initfeh/eep that was never read, so a premature evolutionarymodel: block materialized three free, likelihood-free dimensions (review 3.8.2);
• orbit.fitvcve's WIP guard named the unconsumed mask field as its real blocker, ahead of the missing physics.

The four element roles

role	value from	in the build DAG	potentials	reported
sampled	a raw coordinate of its own	n/a	prior + logit correction	yes
derived	an expression the model CONSUMES	yes	Gaussian + soft barrier	yes
reported	an expression nothing consumes	no (deferred pass)	none	yes
inactive	a bookkeeping pin	no	none	no

Vocabulary, all on the manifest entry: expr_key may be {block: selector} per element; output_expr_key the same for role 3; mask is the activity selector; inactive_value states what a masked-out element is held at (a linear-law band's u2 is exactly 0). One selector normalizer, sized from the parameter's element count -- never the config list (review 1.1.1's hazard) -- raising on a mismatch.

Role 3 is declared, not built. Reporting secosw on a V_c/V_e orbit is derived-but-consumed-by-nothing: ecc reads secosw on the sqrt(e)cos/sin orbits while secosw reads ecc on the V_c/V_e ones -- acyclic per element, a cycle in the per-parameter order graph.py sorts. add_parameter raises a NotImplementedError naming the deferred pass that fixes it, and it lands with its first consumer, the vcve parameterization (8.8.3). It matters because an inactive element's prior is dropped: a coordinate with a computable inverse must flip to reported so a user's prior survives toggling the switch.

Notes worth reading in the diff

• Mixed vectors assemble with pt.set_subtensor, never a pt.where over the two value vectors -- an unused element's pin can legitimately make the expression NaN, and 0*NaN poisons the gradient of the whole vector on every backend.
• Better still, the unused elements never enter the expression: dependencies are sliced to the mask, but only where the alignment is provable (a local param of equal length, a map with one entry per element, or a context node the component declares). A bare cross-component vector raises -- the same call add_parameter already makes for a dep naming a map it does not have.
• "Elementwise" is a property of the function, so both graphs are kept and compared at the start point (System.verify_element_slices): dummy inputs could agree by accident, and evaluating a random variable would draw from its prior instead of reading the start.
• A parameterization flip carries almost every constraint for free -- mu/sigma still apply to a now-derived element, an initval is still a RANK_USER assignment the engine propagates -- and only lower/upper change meaning. inactive is the one lossy case and warns per element, keyed on what the user wrote.
• components/parameterization.py makes declaring a parameterization a table (mode_manifest), with pin_unselected as the separate opt-in pin (through "overrides", so a user who wants the parameter back still wins) -- retiring the three byte-identical pin loops (reviews 4.5.2, 4.5.3).

Verification

• All 19 shipped example configs bit-identical after every commit: start logp compared as the exact float repr, plus the free-RV set, the raw-coordinate count and every parameter's resolved values/bounds/sigma/mu.
• Full suite green (2353 passed), +30 tests.
• Mixed-law and mixed-mass systems build end to end with finite logp and gradient; a mixed model really samples under nuts_sampler="numpyro" (a finite C-backend gradient is not proof).
• A short real fit (kelt4 RV-only) ran the whole pipeline: sampling, modes, LaTeX table, results.csv, the compiled paper draft, plots and the restart file.
• The relaxation engine's per-element seed routing measured, not assumed -- a mass seed reaches whichever coordinate each planet samples, and no engine change was needed.

Also fixed: review 1.2.3 (a pinned vector with a broadcast scalar initval emitted one macro for a body citing several -- an "Undefined control sequence" at the end of a long fit). Closes 3.8.2, 4.5.2, 4.5.3; partially addresses 2.1.4.

🤖 Generated with Claude Code