chore: organize tests and CI by subsystem #113

The test suite is healthy but increasingly difficult to navigate: 23 Python test modules contain about 5,811 lines in one directory, with test_session_peer.py around 1,228 lines and test_codex.py around 764 lines. CI selects some optional suites by manually listing module names while the base matrix uses broad discovery. This makes ownership and coverage boundaries harder to see as relay, MCP, adapters and packaging expand.

Reorganize tests and CI by subsystem without changing product behavior.

Proposed layout

tests/
  core/
  adapters/
  codex/
  relay/
  control_integration/
  mcp/
  packaging/
  docs/
  fixtures/

The exact names may change, but each suite must have one clear owner and invocation contract.

Scope

• Split oversized test modules along existing behavioral boundaries.
• Move relay prototype assertions retained by chore: separate relay prototype history, deployment examples, and operator docs #111 into the product relay suites rather than creating a permanent prototype group.
• Replace fragile hand-maintained CI module lists with documented suite entry points or discovery roots.
• Keep fast core tests separate from dependency-bearing relay/MCP/control integration tests.
• Add explicit distribution-content checks for intended wheel/sdist files and excluded prototype/secret/state artifacts.
• Preserve fixture privacy and deterministic no-network defaults.

Acceptance criteria

• python -m unittest discover -s tests still provides a useful complete local default with documented optional-dependency behavior.
• CI jobs map directly to named directories or stable suite commands.
• No test is silently dropped; before/after test inventory is recorded in the PR.
• Existing OS and Python/Node matrices continue to pass.
• Real-network, OAuth, model and production-service tests remain opt-in and cannot run from default discovery.
• Package tests verify the complete intended artifact contract rather than only a few sentinel files.
• Documentation translation checks remain fast and independently runnable.

Non-goals

• Product source modularization (refactor: modularize the core while preserving the standalone SSH artifact #112).
• Live deployment changes.
• Increasing matrix size solely as part of file moves.

This can follow #111 and may proceed before #112 if it avoids renaming source-facing fixtures twice.