docs: design the micro-compiler architecture for the compilers

[OmarAlJarrah]

Summary

The OpenAPI compiler is one Go package whose files are a filing convention rather than a boundary,
and a single struct has accumulated every lowering concern. The per-file split did not reduce that;
the method count grew. The individual functions are not the problem — no body in the package exceeds
50 lines against a 70-line cap. The failure is type surface, which no rule in force measures.

This adds two documents. docs/micro-compiler-design.md is the architecture: every lowering step
becomes a function over an explicit position, returning its diagnostics instead of accumulating
them, with type interning as the single shared effect, and concerns become packages so the compiler
enforces the boundary. docs/micro-compiler-plan.md is the order the work lands in and what each
step waits on; the detail for each unit lives in its issue rather than being duplicated, and the
blocking relationships are recorded as issue dependencies so the next available piece of work can be
queried rather than inferred.

Three findings surfaced while writing them, each verified against the tree rather than inferred:

• The canonical naming grammar is implemented three times across the OpenAPI, GraphQL and Protobuf
  compilers. Two copies are identical; the third treats . as a word separator and the other two do
  not. Naming.Canonical is part of the IR contract, so the same field means different things
  depending on which compiler produced it, and irverify checks only casing — never segmentation —
  so both spellings pass.
• The architecture test's prefix matching treats an allowlist entry for compilers as licence for
  compilers/graphql, so one compiler may import another today with nothing to catch it.
• The general two-order diff the testing rules prescribe does not exist. It is present as three
  hand-written cases at the site where the pointer-collision bug was found, and the determinism check
  that resembles its general form recompiles the same bytes, which is a different property.
  Separately, nothing anywhere asserts that two distinct source constructs cannot mint the same type
  ID.

Each is dispositioned rather than noted: the first two are prerequisites, and the naming divergence
is filed on its own since it is a live contract violation regardless of whether this work proceeds.

The documents are tracked under docs/ rather than kept as working notes because the previous design
for this area was written to an ignored path, could not be committed, and had to be recovered from
the history of a since-deleted file.

Test plan

Documentation only — no Go files are touched, so no code path changes and the gate is unaffected:

git diff --name-only origin/main...HEAD | grep -c '\.go$' # 0

Every derivation command in the design document was executed against a095636 and its output
matches the surrounding claim; the documents give the command rather than the number wherever a
count would otherwise rot. Every issue referenced in the plan was checked to resolve to the issue it
is described as.