[DRAFT] Run conformance tests as one whole-suite invocation per module

[pedjaradenkovic]

Summary

Iteration 2 of the conformance-testing improvement plan (see CONFORMANCE_TESTING_IMPROVEMENT_PLAN.md on this branch): conformance tests execute as one script invocation per module instead of one per functionality, with shared setup files allowed at the module suite root. Per-FRID subfolders and the test-script contract ("run all tests under $2") are preserved.

Stacked on #252 (Iteration 1); pairs with Codeplain-ai/plain2code_rest_api#123 (backend). Deployment order: backend first (its changes are gated by request flags/fields, so old clients are unaffected).

Changes

• Loop-capable runners (test_scripts/run_conformance_tests_{golang,cypress}.{sh,ps1}): handle $2 being a folder of suites — degenerate single-suite case preserves today's behavior byte-for-byte; loop case runs each non-hidden suite subfolder with === conformance suite: <name> === headers, reports all failures, exits with the first failure's code. Cypress's npm install/build/app-start now runs once per invocation instead of once per suite.
• Whole-suite execution: RunConformanceTests passes the module suite folder (own module) or the required module's copied suite root. The current-FRID run covers all prior FRIDs, so the regression walk collapses to one run per required module (plus an own-suite re-run when code changed during fixing). ~100 lines of FRID-walk machinery removed.
• Failure attribution (render_machine/failure_attribution.py): failed runs are attributed to FRIDs by matching suite folder names against the output; the running context is pointed at the earliest implicated FRID, so the existing fix loop, memory keying, and conflict detection work unchanged. Failure evidence sent to the fix flow is scoped to that FRID (other implicated FRIDs appear as a summary note only, preserving the multi-FRID diagnostic signal). A conservative migration guard fails fast with a regeneration hint on layout-level failures (pre-flip projects, non-recursive custom scripts).
• Shared setup + two-tier guard: render output paths are module-root-relative (tests in the FRID's subfolder, shared helpers allowed at the root); responses are validated before storing — files in other functionalities' subfolders are rejected (one retry, then error), existing shared root files may only be extended by inserting lines (difflib opcode check).
• Summarization step dropped: the per-FRID summarize_finished_conformance_tests LLM call is removed — its only consumer was the plan-stage dedup section, which now receives the actual test files (A/B renders showed identical dedup quality). Saves one LLM call per functionality. The folder_name map in conformance_tests.json stays (attribution and fix routing depend on it).

Testing

• 322 unit tests green (24 new across attribution, folder resolution, and the two-tier guard) + black/isort/flake8/mypy.
• Runner scripts verified across degenerate/multi-suite/hidden-folder/empty/missing layouts (golang with real Go, cypress with stubbed npm/npx; .ps1 variants by review).
• Five e2e renders against a local backend: side-by-side per-FRID vs whole-module verdict agreement; happy path (2 invocations for a 2-FRID render, previously 3 and growing quadratically); a forced cross-FRID conflict correctly attributed to the earlier FRID, fixed-scoped, classified as conflicting requirements, and reported; module-root layout and summaries A/B on a 3-FRID project.