v0.11.5

No rule changed and no finding moved. All 631 findings across the
127-fixture corpus are byte-identical to 0.11.4, paths included. This release
changes how the element rules reach their subject, not what they report, so a
tag that validated a particular way on 0.11.4 validates identically here.

Changed

• Element rules are dispatched by name instead of by hand-written traversal.
  Twice now the same defect has shipped: the CTV Ad Portfolio moved
  <MediaFiles> somewhere new, the rules that check its contents did not
  follow, and a defective <MediaFile> validated clean until someone wrote
  another traversal. 0.11.1 fixed it for <NonLinear>, 0.11.4 for the VAST 2.0
  extension container. A third location would have been a third fix.

  rules/walk.rs visits every element belonging to the document proper, paired
  with the location to report it at. It descends into standardised IAB extension
  containers, whose children are real VAST, and stops at any other
  <Extension> or <CreativeExtension>, whose payload is a vendor's private
  format. That allowlist is now the entire maintenance surface for a new
  container: one entry, rather than one traversal per rule module.

  rules/elements.rs registers the position-independent rules against the
  element name. <MediaFile>, <Mezzanine> and <InteractiveCreativeFile>
  move first, collapsing three call sites into one dispatch.

  Position-dependent rules keep their own traversals, which is correct: their
  subject really is a location. SIMID-1.0-simid-mediafile-required stays
  Linear-only, and the ctv_portfolio.rs rules are about the container rather
  than any one child. The two version constraints survive intact and are
  tested: <Mezzanine> attributes stay 4.1+, and the NonLinear content model
  stays 4.x so a 3.0 document still reports the construct once as an unknown
  child rather than twice.

Performance

Measured against 0.11.4 with criterion, same bench file on both revisions.

The pods and the CTV Ad Portfolio fixtures are at parity (no change detected,
p > 0.05). Small fixtures are within 0-3%. valid_4.3_with_verification is the
outlier at roughly +10%, about 1.3 µs on a 16 µs validation.

The cost lands on documents that gain nothing from the dispatch, which is worth
stating plainly. Where a portfolio container exists the walker replaces the
hand-written traversals it deletes, so the work nets out; where one does not,
the walk is pure addition. It is largest on small, element-dense tags, where a
fixed per-node cost is a bigger fraction of a smaller total.

An eager first implementation built a path string for every visited node and
cost +16% to +19% across the board. Paths are now joined on demand, and sibling
counting is limited to elements that carry an index.

Added

• Benchmark cases for the CTV Ad Portfolio path (ctv_portfolio_9_ads,
  ctv_portfolio_clean) and for element-count scaling (pod_50_ads). None of
  the previous cases contained a standardised extension container, so the
  benchmark could not see the cost of the code that walks one.
• the_same_defective_media_file_reports_identically_everywhere: the same bare
  <MediaFile> placed in every container the spec has put one in must report
  the same rules. This test fails on 0.11.0 and again on 0.11.3, so it would
  have caught both original bugs. Verified non-vacuous by mutation.