Research: split native AMP evaluation from compiled matrix transport

[mountain]

Context

Issue #152 established an exact polynomial-like / matrix-like compiler for

$$ f(x)=x^d+t $$

on a power-dominant chart. Subsequent red-team analysis exposed a solver-layer conflation:

• the Process Geometry observer contract is not a Koopman observable;
• the finite $q$-jet and substitution matrix are valid coefficient-readout presentations, but they are not necessarily the native numerical evaluator;
• expanding $\log(1+tq^d)$ into a $q$-series reintroduces a Taylor-like observer choice.

The exact process recurrence can instead retain exp and log1p as native atoms:

$$ q_{n+1}=\frac{q_n^d}{1+tq_n^d}, \qquad H(y)=y+\sum_{n\ge0}d^{-(n+1)}\log(1+tq_n^d). $$

For $q_0\le1$, the remaining analytic tail after $R$ levels has an explicit bound.

Goal

Split the research-local solver into three typed roles:

1. native process evaluator — default certified-tail numerical path, no Taylor coefficients or matrix construction;
2. compiled polynomial-like evaluator — optional fixed-chart, evaluate-many specialization using structured Horner evaluation;
3. matrix-like transport — exact offline compiler/replay representation, not runtime ontology.

Acceptance

• implement a no-series native process evaluator with declared domain, tolerance, budget, tail bound, and cost ledger;
• prove/test the recurrence and tail certificate against an independent strong recurrence;
• optimize compiled coordinate evaluation using its degree ray;
• keep exact coefficient and residual certification intact;
• separate native, compilation, and online costs;
• include chart-boundary red teams and typed failures;
• record same-accuracy microbenchmarks without promoting them to performance theorems;
• update the Sonnet disposition and Engineering Architecture relation.

Claim ceiling

This issue does not introduce a generic germ, jet, Koopman, matrix-like, or process-evaluator API. It does not claim a universal speedup. The maximum result is a problem-local solver split with measured workload regimes and explicit failure semantics.