research: add native-method lowering firewall

[mountain]

Closes #156.

Summary

Add a research-local, executable method contract that keeps four roles distinct:

• native discovery;
• native evaluation;
• certificate;
• conventional baseline / red team.

Classical mechanisms are not banned. They may run in a declared baseline or certificate lane. They may enter a native lane only through a task- and lane-scoped LoweringWitness that records adequacy, preserved and forgotten information, residual, decoder, certificate, and failure semantics.

The trace also rejects attempts to relabel baseline evidence as native evidence and emits a machine-readable multi-axis cost audit.

Evidence

• 23 exact firewall tests;
• AMP native-evaluator and repository semantic/documentation hygiene integration;
• full suite: 796 passed, 25 skipped;
• git diff --check clean.

Mathematical Core relation

Implements a research-method audit around the existing chain

history -> task -> presentation -> retained fibre/residual -> decoder.

It adds no mathematical object or law.

Engineering Architecture Change

• Stage: research problem/task intake, evaluator selection, certification, and baseline audit.
• Object/task: caller-declared research method records; no generic solver carrier.
• Algorithm/backend: exact Python enum/dataclass validation and append-only trace; no new dependency.
• Claim mode: exact finite validation of the declared record, not proof that arbitrary source code was honestly labelled.
• Failure semantics: incomplete contract, unknown mechanism/lane/task, premature lowering, undeclared baseline, and evidence-lane mismatch fail closed.
• Residual/decoder: required on every lowering witness.
• Cost: separate discovery, compilation, evaluation, live-state, stored-history, residual, and decoder counts; scalarization is unauthorized.
• API/dependency effect: none; files remain in workstreams/.
• Operation: support/refine the existing Effective Analysis discipline; no architecture promotion.

Theory Map Change

Unchanged. This is T0/T1 research tooling, not a universal calculus, canonical solver, ontology, Experimental abstraction, or Public API commitment.

Red team and boundary

Typed mechanism names prevent accidental aliasing but do not sandbox Python or prove source honesty. Independent mathematical review and executable certificates remain necessary. The first downstream calibration will be a Brownian scale/fibre Sonnet that keeps Gaussian, heat-kernel, Fourier, and PDE machinery outside blind native scale discovery.