fix(internal/harness): exclude permutations that change meaning

[OmarAlJarrah]

Summary

The two-order oracle compiles a source as written and again with every mapping
reversed, and reports what the permutation changed. Its correctness rests on the
rewrite being meaning-preserving. reverseMappings documented one case where it
is not — a duplicate mapping key, whose meaning depends on the order being
changed — and excluded it. Three more went unexcluded, and in each the two
compiles differ because of the rewrite rather than because of a lowering, so the
oracle reported a defect in the compiler that was not there.

• A flow-style implicit null is not preserved. yaml.Marshal re-emits {A}
  with an empty string for its value, so the second compile read a string where
  the first read null. Both arms are now compiled from the encoder's output,
  which leaves declaration order as the only difference between them.
• Reversing a mapping can carry an alias above its anchor. YAML requires an
  anchor to be defined before it is referenced, so the permuted source did not
  parse and the parse failure was reported as order dependence. The rewrite now
  re-parses what it produced and declines a source whose permutation does not
  survive it, which covers any later ordering rule of the same kind without
  enumerating it.
• diagnosticSet identified a finding partly by Provenance.Pointer, which
  ir/provenance.go documents as holding either a structural pointer or
  line:col. A permutation moves a construct to a different line by design. A
  positional pointer is now replaced rather than dropped, so a permutation that
  changes how many findings were reported still shows.

All three were latent. No committed spec reaches any of them and the corpus
sweep was green; they surface once a spec using one of these spellings is
committed, or the oracle is driven by generated input.

TestReverseMappings_AliasIsNotFollowed asserted the anchor-above-alias output
as correct — invalid YAML that the oracle was handing to the compiler. Its
fixture moves into a sequence, whose order the rewrite leaves alone, so it still
pins that an anchored mapping reverses once rather than once per alias; the shape
it used to assert is now covered by a test that requires it to be refused.

Closes #241

Test plan

• Each of the three cases is pinned by a test that fails without its own fix and
  passes with it. Reverting the fixes one at a time reddens exactly one case
  each, so no test is standing in for another:
  • baseline normalization → flow-style implicit null
  • re-parsing the permutation → an alias reordered above its anchor and
    TestReverseMappings_AliasAboveItsAnchorIsRefused
  • positional pointers → a diagnostic located by line and column
• The change adds skip paths, and a skip is how this oracle goes quiet without
  the sweep noticing, so reach was measured rather than assumed: 82 corpus specs
  still reach the comparison, and the new baseline skip fires on none of them.
  The seven that are newly declined are the deliberately-invalid anchor and
  duplicate-key fixtures, which stop at an earlier oracle and never reached this
  one.
• TestOrderInvariant_ReachesTheCorpus now guards both arms rather than the
  rewrite alone. Forcing either arm to decline reddens the suite.
• Full gate on the rebased branch: gofmt, go vet, golangci-lint run
  (0 issues), go build ./..., and ./scripts/check-coverage.sh at 100% of
  4373 statements.