feat(ingestion): business-logic analyze phase — populate likely_plumbing + candidate_business

[theagenticguy]

Summary

Makes the merged sieve kernels (#248) end-to-end: codehub analyze now writes likelyPlumbing + candidateBusiness into nodes.payload for every Function / Method / Constructor / Class / Interface / Struct in a Python / Java / Go repo. The user gets both concern tags from two commands (codehub init + codehub analyze) with no query, no labels, no embeddings:

likely_plumbing    — precision-first (~0.94)  "this is plumbing"
candidate_business — recall-first    (~0.93)  "look here for domain logic"

Queryable via SQLite JSON1: SELECT name FROM nodes WHERE payload->>'$.candidateBusiness' = 'true'.

Components

• core-types — two optional CallableShape fields. Auto-persist through nodes.payload; no storage-adapter change (the SQLite store rehydrates payload verbatim).
• extract/business-logic-features.ts — faithful Python→TS port of the feature extractor. Reproduces the marker logic exactly: word-boundary / camelCase-component matching, the precise nPlumbingSignals formula (serialization + observability + getter/setter + dto-mapper-ratio≥0.5), and ORM-base class-head detection. 44 unit tests.
• pipeline/phases/business-logic.ts — the analyze-time phase (after complexity). Slices each symbol body, runs classifyPlumbing + classifyBusinessCandidate, re-adds the node with the tags (richer-entry-wins merge, same contract complexity uses). Python/Java/Go only; other languages skip silently. The class-head slice scans upward over a Javadoc block to reach the real @Entity / @MappedSuperclass annotation while excluding @author-style comment tags.
• default-set + orchestrator test — registered after complexity; the topological-order assertion updated for the new position.

The parity contract

The whole point of the port is that the shipped numbers survive it. An independent per-symbol harness diffs the TS analyze-pass verdicts against the Python oracle across all four corpus repos:

1368 / 1368 = 100.0% verdict agreement, 0 disagreements.

This was re-verified independently of the porting agent's own report — which is how a JPA-entity divergence surfaced (a Javadoc @author tag shadowing the real ORM annotation, dropping isOrmModel and flipping 5 entity classes). Caught at 99.63%, fixed to 100%. The 0.936 plumbing precision / 0.925 business recall hold end-to-end.

Determinism

computePlumbingFeatures and both kernels are pure; files + definitions iterate in sorted order. Tags are byte-stable across runs (the three graphHash-determinism tests pass), safe under the reproducibility contract.

Verification

• core-types / analysis / ingestion typecheck clean.
• ingestion 629/629, core-types 83/83, analysis 14/14.
• biome + banned-strings + commitlint + pre-push hook pass.