Build an interval specification from a description of the analysis

[billdenney]

Closes #157.

pknca_interval_table() takes when the interval runs, how the drug was given,
how often, and what was collected, and returns the interval specification that
describes it:

pknca_interval_table(0, 24, dosing = "single", route = "extravascular")
#>   auclast aucinf.obs aucpext.obs cl.obs cmax tmax count_conc half.life tlag
#>   impute = start_predose_conc0

Context	Parameters	Imputation
single, extravascular	AUClast, AUCinf.obs, %extrap, CL, Cmax, tmax, count, t½, tlag	start_predose_conc0
single, IV bolus	same minus tlag, plus c0 on its own row	start_predose_conc0; c0 gets none
single, IV infusion	plus ceoi	start_predose_conc0
single, continuous infusion	no ceoi	start_predose_conc0
steady state	AUCint.last, AUCint.inf.obs, CL.int.inf.obs, Ctrough, …	start_predose
multiple, not steady state	same	start_cmin
urine	Ae, Fe, CLr, volume	none
sparse	sparse AUClast + SE	none

pknca_presets() gives named argument sets (bioequivalence, first_in_human,
mass_balance, …); explicit arguments override a preset.

Why a specification can have more than one row

For an IV bolus, c0 is calculated without imputation while everything else
uses it. Before the first dose a predose measurement is contamination rather
than the concentration at the dose, and pk.calc.c0() returns an observed
nonzero concentration at the dose time in preference to back-extrapolating. An
end-to-end test asserts c0 from the generated specification equals c0
calculated with no imputation at all.

The imputation pairing rules replace the per-parameter impute_conflict field
the plan originally called for. An experiment across every parameter and every
imputation method (1180 interactions) found that almost none of the conflicts
are context-independent — c0 with a shifted predose is correct at steady
state and wrong for a first dose, and so on. Since this function chooses both
the parameters and the imputation, it can simply not pair them, which handles
the context-dependent cases correctly and needs no new registry field.

Three classification errors this exposed

None were caught by the 50 existing classification checks:

• aucint.inf.obs/aucint.inf.pred were single-dose only. The "inf" names the
  extrapolation used for the tail, not the end of the interval — over a bounded
  interval they give AUCtau. Steady state was left with no interval AUC at all.
• sparse_auc_se and the other diagnostics produced alongside a sparse AUC were
  not flagged sparse in the registry, because pk.nca() only needs the flag to
  route a calculation. They leaked into every dense context. Now classified
  through what they depend on.
• The AUMC parameters had no basis to be built on, because the basis named only
  areas, so the mean residence times and steady-state volumes were unreachable.

Regression tests added for the first two.

Reachability

174 of 207 parameters are reachable across the 96-context grid. The 33 that are
not are the AUCall family, pinned as an exact list: whether AUCall differs from
AUClast depends on values below the limit of quantification after the last
measurable one, which is a property of the data rather than of the analysis, so
no context chooses it. include = "aucall" reaches it, and a test shows that.

The sweep needed a cache to run at all — 96 contexts × 28 dependency-closure
walks timed out. Basis families are cached alongside the classification and
dropped together when the registry changes.

80 new tests; full suite passes.

🤖 Generated with Claude Code

[billdenney]

This is overly nested. Simplify it.

[billdenney]

Remove lambda.z.n.points from here

[billdenney]

Remove the extra c()

[billdenney]

Please move pknca_preset_definitions to just be this function unless there is a clear reason to have two functions.