[jeremiepas]

Why

Running --cluster-algorithm bunch on solario-core, the 57 dataset-discovery repositories and five Solario builds (2026-11-12) needed four workarounds outside Graphos. Each one hides a defect:

#	Defect	Measured effect
1	The community-size cap (5 % of nodes) also applies to Bunch	nestjs-realworld (29 files): 29 one-file modules, MQ 0, instead of 9 modules, MQ 4.20. solario-core src/cli: MQ 16.66 instead of 18.37
2	Step 5 re-clusters on depends_on/inferred edges derived from Step 4, and the MDG also folds documents edges	src/infrastructure: 1,490 inferred edges against 316 real imports
3	A search is intractable above ~600 files (> 28 min at 1,778), but the Leiden fallback only triggers at 5,000	No Bunch result for any real repository above ~600 files
4	resolveImport never sees the corpus	Extension-less paths, @/ ~/ $lib/ aliases, tsconfig paths and JVM FQNs stay unresolved. taxonomy: 0 file-to-file imports; jhipster: −224 internal imports

What

openspec/changes/bunch-real-projects/ contains the proposal, the design (D1–D4, with a Mermaid flow), the tasks, and spec deltas for bunch-clustering, pipeline-clustering and import-resolution.

• D1. No community-size cap on a Bunch partition.
• D2.
  • The MDG uses a relation allowlist: imports, calls, references, extends, implements.
  • Step 5 reuses the Step-4 Bunch partition instead of searching again.
• D3. Partitioned search over directory sub-trees (≤ 500 files), then a global refinement pass:
  • Bunch's single-file moves over the whole MDG, every import counted.
  • MQ strictly increases with every move, so the pass terminates.
  • Validated externally: pc26 (3,929 files) went from 1,579 singletons and MQ 149 to 1 singleton and MQ 376.
• D4. Corpus-aware import resolution, with a uniqueness rule. Ambiguous targets stay external.

--edge-density sparse is left to the open cli-enum-flag-vocabulary change.

Verification

• openspec validate bunch-real-projects --strict: valid.
• Spec and design only; the implementation is in tasks.md §1–5.
• The validation targets in design.md reproduce the numbers measured with the external workarounds.