Make the startup collector summary honest about disabled collectors

Description

Verified live on the shipped v2.0.1 build. With an operator explicitly setting COLLECT_INSIGHT=false, the startup banner logs that setting as a warning line and then two lines later reports skipped_collectors: 0, lists insight in both configured_names and ‘Enabled Collectors’, prints a cadence for it under ‘Collector Cadences’, and starts a per-collector loop task for it (task_count 9). Reading the startup logs, an operator concludes the collector is on.

The runtime behaviour is correct: collectors/insight.py:178 returns an empty endpoint-group tuple and collectors/insight.py:277 returns early, so no API call is made and no error series appears. The disagreement is between the log summary and reality, and it is the same shape as MDE-0065/MDE-0066 - a top-level surface reporting health that the internals contradict.

The mechanism is that the collector-level is_active property (core/collector.py:79, overridden at collectors/insight.py:189 and collectors/clients.py:128) is already the intended signal and is consulted by the web UI (app.py:1125) and the manual-trigger control API (app.py:1441), but not by the startup summary in app.py:1161-1190 or by the manager’s registration and loop-start path in collectors/manager.py:262-330. collectors.active_collectors is a separate setting from the per-collector collect_* flags, so a collector switched off through its own flag is never added to skipped_collectors.

Scope is the log summary and the no-op loop, not the collector gating, which already works.

Acceptance Criteria

• #1 The startup collector summary counts a collector disabled by its own collect_* flag as skipped, with the flag named as the reason, rather than reporting it enabled
• #2 ‘Enabled Collectors’ and the printed cadences exclude a collector whose is_active is false
• #3 No per-collector loop task is started for a collector whose is_active is false
• #4 A regression covers a disabled-by-flag collector and asserts the summary, the enabled list and the started task set all agree with is_active

Definition of Done

• #1 just check (ruff format –check, ruff check, mypy, generated-doc drift, offline API conformance, and the marker-filtered pytest run with the 80% coverage floor — this is exactly what the CI test job runs)
• #2 just gen, when metrics, config, endpoints, collectors, the settings schema or the chart config changed — just check includes the drift gate and CI fails the build on it
• #3 Grafana queries in grafana/dashboards/*.json and grafana/alerts/ updated, if a metric or label name changed

Implementation Plan

1. Failing tests first in tests/unit/test_disabled_collector_reporting.py: a flag-disabled collector must be absent from CollectorManager.collectors, present in skipped_collectors with the flag named, still addressable through get_collector_by_name so the control API keeps answering ‘disabled’ rather than ‘not found’, absent from get_scheduling_diagnostics()[‘collectors’], absent from the startup summary’s Enabled Collectors, and given no per-collector loop task.
2. Single root fix in collectors/manager.py _initialize_collectors: after instantiation, when collector.is_active is false, register the instance’s metadata (index + lock) but route it to skipped_collectors instead of self.collectors, with a reason naming the disabling flag. Every downstream surface reads self.collectors, so the summary count, cadence gauges, scheduling diagnostics, readiness set, collect_initial ordering and app.py loop-start all become correct from that one change.
3. core/config_logger.py log_startup_summary takes the live active and disabled collector names instead of reading settings.collectors.active_collectors, which cannot see a per-collector flag. app.py _log_startup_summary supplies them.
4. Gate: just check, plus a CodeRabbit pass since this is code with branching.

Implementation Notes

Verified failing-first: with collect_insight=false the five new regressions failed while the captured startup log reproduced the live defect verbatim - Enabled Collectors carried ‘insight’, Collector Cadences printed InsightCollector@900.0s, and task_count was 9 including InsightCollector.

Root fix is one branch in collectors/manager.py _initialize_collectors. An instantiated collector whose is_active is false is registered in the name index and lock map (so the control API still answers ‘disabled’ rather than ‘not found’), appended to skipped_collectors with its own flag named, and never appended to self.collectors. Every downstream surface already reads self.collectors, so the summary count, cadence gauges, scheduling diagnostics, readiness set, collect_initial ordering and app.py loop-start all corrected from that single change. A new MetricCollector.inactive_reason property carries the operator-facing reason; InsightCollector and ClientsCollector override it to name collectors.collect_insight and clients.enabled.

log_startup_summary now takes active_collector_names and disabled_collector_names, because settings.collectors.active_collectors is only the name allow/deny list and cannot see a per-collector flag. It falls back to the old source when the caller supplies neither, and prints a new Disabled Collectors line.

Both collectors this touches are off by default, so a default deployment now reports clients and insight as skipped instead of enabled. That is the intended correction, and no existing test depended on the old behaviour.

Gate: just check exit 0 - ruff ‘All checks passed!’, mypy ‘Success: no issues found in 122 source files’, 2965 passed / 5 deselected, coverage 91.51%. CodeRabbit review completed with 0 findings across all 8 changed files. just gen was not needed: no metric, config, endpoint, collector-registry, settings-schema or chart input changed, and the drift gate inside just check passed. No Grafana change: no metric or label name moved.

Final Summary

Shipped in a231a50. A collector switched off by its own collect_* flag now reports as skipped with that flag named, instead of being counted as enabled, printed under Enabled Collectors with a cadence, and handed a per-collector loop task. Proved by eight new regressions in tests/unit/test_disabled_collector_reporting.py, five of which failed first against the live defect, and by just check green at 2965 passed / 91.51% with a clean CodeRabbit pass.