feat(architecture): show a repository's real modules and what they define

[parthrohit22]

Closes #445.

What was wrong

The engine was already right; the view discarded it. On pallets/click — 177 files, 1,963 observed symbols, truthClass: observed — the Architecture view rendered the entire library as one node:

{ "id": "module:Click", "name": "Click",
  "description": "Click derived from repository intelligence at /src/click.",
  "responsibilities": ["Owns utility concerns"] }

What it does now

A file that defines symbols is a module. The old rule grouped everything under src/<package>/ into one id, so a single-package repository became a single box. Files that define nothing — documentation, tests, config — keep their existing role grouping, so nothing else in the view moves.

Descriptions state observed facts. Symbol count plus the symbols a reader would recognise the module by. Every clause traces to a fact already sealed in the snapshot; nothing infers what a module is for.

Responsibilities stop asserting ignorance. "Owns unknown concerns" was a statement about the classifier wearing the costume of a statement about the code. Where a role was classified it is reported; where it was not, the entry is omitted.

Colliding names get qualified. FastAPI has four modules called utils; they now read openapi/utils, security/utils, and so on. Unique names stay short.

Verified live

pallets/click:

module	description
core	Defines 38 symbols, including Argument, Command, Context, Group.
exceptions	Defines 22 symbols, including Abort, BadArgumentUsage, …
decorators	Defines 18 symbols, including argument, command, group, …
parser	Defines 17 symbols, …

That ranking is the library as anyone who knows it would describe it. fastapi: 3,139 files → 358 nodes (the documented bound holds), and applications reads "Defines 1 symbol: FastAPI."

Design notes

Symbols reach the analyzer through a key-only query. architecture_facts deliberately excludes symbol rows because they dominate a large snapshot. Stable keys are short strings and already carry <path>::<qualified> — all the module inventory reads — so the bound that matters is preserved.

Only top-level definitions count. A method is defined by its class, not by the module. Counting every method turns "what does this module define" into a line-count proxy, which is the synthesized measure #217 rules out.

One acceptance criterion deliberately not met

not_computed fields are either populated or removed from the response.

I left estimatedLines / estimatedComplexity as they are, and think they should stay:

• They cannot be populated honestly. The only line count in the store is RiEvidence.logical_line_count, which is documented as internal validation metadata — producer-supplied, deliberately excluded from the canonical evidence record and the graph hash. Two equivalent graphs may carry different values. Promoting it to a displayed number would be exactly the unsourced claim this product refuses to make.
• Removing them loses a signal. An explicit not_computed is the same honest-limits posture as not_assessed elsewhere: it distinguishes "nothing measures this" from "the field is missing." The UI already renders nothing for not_computed (ArchitectureNode.tsx:128), so there is no user-visible noise — only JSON.

Happy to strip them in a follow-up if you'd rather; it touches the schema, the report builders, layout.ts and the generated contract, so it doesn't belong in this PR either way.

Tests

Full backend suite green (pytest -q, 13 skipped, 0 failed); ruff check and ruff format --check clean. tests/test_architecture_relationships.py fixture ids updated to the new per-file module ids. Response shape is unchanged, so the generated frontend contract needs no regeneration.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
partha-frontend	Ready	Preview	Sep 11, 2026 6:35pm UTC