fix: keep exports of a required module listed before its dependent

[bananaplain]

A module whose requires lists an ancestor and then a module that itself requires that ancestor (requires: [a, b], with b requiring a) failed to parse when b uses a concept exported by a:

Syntax error: Concept :A: is not defined in the definitions.

process_required_modules skips a the second time it is reached, so it isn't processed twice. That skip also dropped a's exported definitions, so b was validated without them. Every module above b in the chain failed with the same error.

The skip comes from codeplain, where it is harmless: codeplain validates each required module in a separate, fresh parse, where b processes a itself. Since 0.1.1, plain-parser validates required modules inside the top module's parse, so that plain-parser check matches --dry-run while reading each file once. In that shared walk, the skip loses a's exports.

Change

• Each module's exported definitions are recorded the first time it is processed and returned again when the module is skipped. Only its own exports, never its ancestors'.
• No condition changes: diamonds and independent branches are rejected as before. Two shapes now fail with the fixed-order error where 0.1.2 reported a missing concept, because b no longer fails first: requires: [b, c, b], and requires: [base, left, right] with left and right both requiring base. That is the message codeplain gives.
• One shape is newly rejected, as codeplain rejects it: b defining a concept with the same name as one a exports, with different text, fails with "Concepts were defined multiple times".
• New fixtures tests/data/requires/shared_export_{base,middle,top,above}.plain cover the module that requires both (top) and a module above it (above).
• Topology tests: accepted chains, rejected diamonds and independent branches, and two valid chains still rejected (as in codeplain), marked as expected to fail. The cases live in named constants, and the tests around the shared base are named after the modules they parse.

Verification

• Tests, black, isort, flake8 and mypy pass.
• plain-parser check across 3,725 example and corpus .plain files: compared with 0.1.2, the only changes are three existing modules with this requires shape, which go from failing to OK, plus the new fixtures.

Release

• Ships as 0.1.3.
• plain-forge users on 0.1.2 need uv tool upgrade plain-parser; plain-forge update does not upgrade it.
• Known difference from codeplain, left for 0.1.4: when more than one module in a chain has an error, plain-parser reports the deepest module's and codeplain the top module's. Accept/reject and the marshalled trees already match.

[zanjonke]

LGTM 🚀

[hisenb3rg]

👍