start method

@override FutureOr<void> start()

Loads plugin configuration (enabled rules, severity overrides, etc.) from analysis_options / SAROPA env vars before rules are registered.

Logs the start event via PluginLogger — the entry will be buffered in memory until loadNativePluginConfigFromProjectRoot runs and PluginLogger.setProjectRoot is called on the first analyzed file, then flushed to reports/.saropa_lints/plugin.log so users have a visible surface confirming the plugin actually started.

Implementation

@override
FutureOr<void> start() {
  PluginLogger.log('Plugin.start() — loading initial config');

  // Arm the rapid-edit gate: this runs ONLY inside the interactive analysis
  // server (no bin/ CLI instantiates the plugin), so it is the reliable signal
  // that in-flux relief is safe. Batch runners (scan/baseline/health, `dart
  // analyze`) leave this false and therefore report every rule at full
  // fidelity. See SaropaLintRule.deferForRapidEdit / isAnalysisServer.
  SaropaLintRule.isAnalysisServer = true;

  try {
    loadNativePluginConfig();
  } on Object catch (e, st) {
    PluginLogger.log(
      'loadNativePluginConfig failed in Plugin.start()',
      error: e,
      stackTrace: st,
    );
    // Defensive: plugin still registers with defaults
  }

  // Arm the memory-relief subsystem. Before this call,
  // initializeCacheManagement() was defined but never invoked anywhere, so
  // MemoryPressureHandler stayed disabled: none of the plugin's own caches
  // (compilation-unit, file-content, metrics, source-location, semantic
  // tokens, import-graph, string interner, …) were ever registered for
  // eviction, and auto-relief never armed. They therefore grew for the
  // entire analysis-server process lifetime. Wiring it here bounds the
  // plugin's OWN footprint and sheds it under pressure.
  //
  // Scope honesty: this caps the plugin's caches only (sub-GB on a large
  // project). It does NOT bound the analyzer's resolved element/AST model,
  // which is the dominant cost when many element-resolving rules run over a
  // large codebase under strict modes — that is reduced by doing less
  // resolution, not by cache relief.
  try {
    initializeCacheManagement();
    // initializeCacheManagement registers the project_context caches but not
    // the report-layer ImportGraphTracker, which holds a per-file set of
    // import/export URIs for every analyzed file and is never evicted across
    // the server lifetime. Register it for pressure relief at clear-late
    // priority (rebuilding the graph requires re-walking files, so shed it
    // only after cheaper caches). Non-destructive in normal operation — it
    // is cleared only when the memory estimate crosses the threshold, and
    // repopulates as files are re-analyzed.
    MemoryPressureHandler.registerCache(
      'importGraphTracker',
      ImportGraphTracker.reset,
      priority: 85,
    );
  } on Object catch (e, st) {
    PluginLogger.log(
      'initializeCacheManagement failed in Plugin.start()',
      error: e,
      stackTrace: st,
    );
  }
}