v2.1.10

Assurance: hostile-input fuzz CI gate + per-parser resource-bound tests

Promotes the hostile-input fuzz suites from incidental coverage under the
generic test job to a dedicated, separately-visible CI gate, and locks the named
resource caps at every parser boundary with oversized-input assertions (the
roadmap "Resilient" track: proven resource bounds, fuzz promoted to a gate).

• New hostile_input pytest marker and a dedicated hostile-input-fuzz CI job
  that runs the marked tests at a higher Hypothesis example budget (a ci-fuzz
  profile, loaded when RECON_FUZZ is set). The render and Bayesian fuzz suites
  and the resilience-hardening tests carry the marker; they still run inside the
  normal test job where they count toward coverage, so a parser-bound regression
  is now its own red check.
• tests/test_hostile_input_bounds.py drives crafted oversized / flooded /
  malformed input straight at each parser and asserts the cap holds: userrealm
  _MAX_AUTODISCOVER_DOMAINS; crt.sh _MAX_SANS_PER_CERT and
  _MAX_CRTSH_CERT_SUMMARY_ENTRIES; the CT burst and wildcard-cluster caps; the
  SPF redirect depth bound; and the DMARC rua extraction under a mailto flood.
• The Autodiscover XML parser now degrades cleanly on a defusedxml
  entity-expansion (billion-laughs) or external-entity (XXE) payload: those
  raise EntitiesForbidden / ExternalReferenceForbidden (not ParseError),
  which previously propagated out of the parser; the guard now catches the
  defusedxml base exception and returns an empty result, asserted by the gate.

Gate: full pytest (2867 passed), ruff, pyright (0 errors), validate_fingerprint (841), branch coverage 85%.