docs(discovery): add the ADR Discovery module specification

[pengyuzhang]

What this adds

One file: Discovery/adr_discovery/README.md, the specification for the ADR Discovery endpoint collector.

Discovery already has a test plan in the repository (Discovery/tests/README.md) but nothing describing what the collector is, what it has to find, or how it is built. This fills that gap.

What it covers

• The four targets — AI binaries, AI agents, MCP servers and skills — with where each lives and what counts as proof, plus the two categories excluded on purpose (instruction files, and the scripts and schedulers that start an agent).
• The six sources the enumerator queries before it searches: package databases, application registries, the kernel, application state, network egress, and the platform execution journal.
• The M1–M7 module split with its three cross-cutting concerns — catalog, redaction and coverage.
• The code layout that split implies, and the import rules that keep it: one directory per module, only world/ touching the OS, no stage importing a sibling, all checkable by a single walk of the import graph.
• The asset and coverage shapes, the redaction rules, the evaluation layers, and the build order.

Notes for review

• The status section says plainly that the implementation is not in this repository yet, so the file cannot be misread as documenting code you can import today.
• Two categories are excluded from discovery deliberately, and both are named in coverage.out_of_scope rather than silently omitted — so a clean machine stays distinguishable from an unasked question.
• Accuracy claims state which measurement layer produced them. A fixture suite built from the catalog measures internal consistency and is labelled as such; it is not reported as precision or recall.