fix: match real GitLab event action_name values

Andrew Dunnrequested to merge
fix/event-action-matching into main

Problem

GitLab's Events API (GET /api/v4/users/:id/events) returns human-readable action_name strings, not clean tokens. A comment is "commented on"; a push is "pushed to" or "pushed new" depending on ref state. The analyzers matched the raw strings literally:

  • internal/analyzer/collaboration.go: e.Action == "commented" -> on a real instance the comment term is always zero.
  • internal/analyzer/source_control.go: e.Action == "pushed" -> push counts collapse on a real instance.

The fixtures hardcoded the clean tokens ("commented", "pushed"), so every test passed without ever exercising a real value. The bug shipped green.

Fix

Normalize once at the source. New normalizeAction helper in internal/collector/events.go, called from toSchemaEvent, maps raw action_name to a canonical token:

  • prefix pushed -> "pushed" (covers "pushed to", "pushed new", "pushed to and removed")
  • prefix commented -> "commented" (covers "commented on")
  • all other actions preserved verbatim, trimmed and lowercased

Analyzers are made resilient as defense-in-depth: both now match the canonical token via strings.HasPrefix. Branch creation in Source Control reads the structured push_data (push_data_action / push_data_ref_type), not the action string, so it survives any wording.

De-vacuify the tests

  • Collector fixtures (events_test.go, integration_test.go) now feed realistic raw values: "commented on", "pushed to", "pushed new".
  • New TestNormalizeAction and TestToSchemaEvent_NormalizesActionName pin the raw-to-canonical mapping.
  • New TestScorers_RealActionNames asserts an event with action_name: "commented on" yields a nonzero collaboration comment signal (computed 1.7) and that "pushed to" / "pushed new" count as pushes with branch creation (computed 1.2).

Validation

  • make check passes (go vet ./... + go test ./...).
  • go test -race ./... passes.
  • Fails-on-old proof: reverting only the normalization and matching fix makes the new tests fail with exactly the zero-signal symptom:
    • TestScorers_RealActionNames: collaboration: got 0.800000, want 1.7 and source control score is zero
    • TestToSchemaEvent_NormalizesActionName: action_name="commented on" -> Action="commented on", want "commented" Restoring the fix makes them pass.

Merge request reports