feat(ml-worker): add disposition corpus export census

[Xore]

Summary

• add a read-only ml-anomalies export/census command with one point-in-time snapshot for status totals and closed dispositions
• redact free-text disposition reasons by default, write atomic NDJSON/JSON outputs with SHA-256 metadata, and report explicit precision-only and calibration gates
• document the alert-population limitation: below-threshold traffic is not persisted, so this corpus cannot measure deployment recall
• add fixture tests for denominators, open/legacy exclusion, class balance, redaction, missing fields, and safe Elasticsearch failures

Validation

• python3 -m pytest ml-worker/tests/test_disposition_corpus.py -q — 11 passed
• python3 -m pytest ml-worker/tests/ -q — 319 passed
• local fixture smoke run — printed the precision-only banner, exported only closed labels, redacted reasons, and verified the report digest
• read-only live Elasticsearch census through the deployed worker (ephemeral snapshot) — reached ES 9.5.3; 7,675,439 alerts, all open, zero labels, time range 2026-09-05T22:00:04.700Z–2026-09-25T07:12:42.167Z

Closes #3295

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None

[Xore]

The two failing checks on this PR are pre-existing breakage on main, not caused by this diff. This branch touches no backend or container files.

• Breaking commit: 8ec35d3e (chore(deps): bump rand (#3287)) bumped rand 0.9 → 0.10 in the backend-service without migrating call sites.
• It merged with only the Dependabot workflow green; the full container CI did not run, so the break landed unnoticed.
• Tracked as main does not compile: rand 0.10 breaking bump landed without call-site migration #3299 with the affected call sites and the two fix options.

Every PR running the backend-service job is red until that is fixed. Not fixed here to keep this diff to its own scope.