fix(ml-worker): quarantine malformed events in the retrain path so one bad row cannot kill the worker (#2984)

[Xore]

Why

Retrain window data is attacker-influenced. #2978 fixed one crash trigger in that
path; this makes the next unknown one survivable instead of fatal, using the
same caller-side guard pattern as #171.

quarantine_retrain_events() probes each row through both
extract_features() and compute_batch_session_features([row]) before the
retrain proceeds, since either entry point can raise on a malformed event.
Dropped rows are counted and surfaced via write_retrain_metric (with an ES
mapping addition) so quarantine is observable, not silent. Non-retrain readers
stay strict — this isolation is scoped to the retrain path only.

Evidence

5 new isolation tests plus a regression covering the honeypot.session list
shape; full ml-worker suite: 298 passed.

Closes #2984

[strix-security]

Strix Security Review

No security issues found.

_{Updated for ceadfe7.}

---

Reviewed by Strix
_{Re-run review · Configure security review settings}

[github-actions]

Dependency Review

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

Scanned Files

None

[strix-security]

Reviewed the full diff of ml-worker/worker.py and its accompanying tests. The PR adds per-row crash isolation for the retrain path via quarantine_retrain_events(), which probes each event's _source through the pre-existing extract_features() and compute_batch_session_features() entry points before retraining, drops rows that raise, and records the dropped count in the retrain metric. This reuses the established quarantine pattern from the live scoring path and introduces no new attacker-reachable sinks, no relaxed authorization, and no handling of untrusted data beyond what the unchanged feature extractors already processed inside retrain(). The new dropped_count ES mapping field is a scoped integer addition. Static analysis surfaced three MD5 usages at lines 294, 326, and 785, all outside the changed hunks and therefore out of scope for this review. No security issues were identified in the changed code.

---

Reviewed by Strix
_{Configure security review settings}