- load_baseline raises SystemExit with regen hint when the JSON file
  is malformed instead of dumping an opaque traceback.
- write_baseline cleans up its tmp sidecar on any exception so a
  crashed run can't poison the next refresh from the same PID.
- The validator now exits 1 (not 0) when AXONFLOW_OPENAPI_SPECS_DIR
  is set but doesn't point at a directory. A misconfigured CI step
  silently disabling the gate produces a green check on a non-running
  validator, which we refuse to do. An unset env still skips with 0.
- Workflow's SHA-bump guard distinguishes "baseline missing on base
  branch" (genuine first-pin introduction) from "baseline present
  but unparseable" (bypass attempt). A malformed baseline on the
  base branch can no longer route a labelless PR through the
  first-pin path.
- Java source parser:
    * Recognises Java 15+ text blocks (`"""..."""`). Previously the
      first `"""` looked like quote-empty-quote and the body parsed
      as live source — fake `@JsonProperty(...)` and `class X {`
      inside a text block silently corrupted attribution.
    * Filters annotation matches whose position falls inside a
      blanked range (string, comment, text block) so the wire-name
      regex on raw content can't grab annotations the structural
      pass already nullified.
    * Attributes record-parameter annotations to the record. Pending
      decls are now tracked between their token and the body `{`,
      so `record Foo(@JsonProperty("x") int x) {}` no longer drops
      the wire name. The SDK has no records today; this is forward
      defence for the next type-class refactor.
- Documents the remaining anonymous-inner-class limitation (no
  `class` keyword, so annotations would leak to the enclosing type).
  The SDK does not put @JsonProperty inside anonymous inners.

Synthetic-fixture tests cover record params, text blocks with fake
annotations, line and block comments, nested classes, and enum
values. Real Java SDK baseline unchanged: 70 registered types, 39
drift, 8 cross-spec, 1 intra-file.

Gate 1 (cross-spec divergence) only iterated currently observed
schemas. A baselined divergence that the platform has since
reconciled silently lingered in the baseline forever; the same old
incompatible shape could be reintroduced and pass the gate because
its fingerprint matched a stale entry that should have been deleted.

Adds the reverse pass that Gate 2 already does for intra-file
duplicates: any baselined cross-spec name that is no longer
observed in the current specs fails the run with a pointer at the
specific baseline key to delete.

Verified locally:
- Positive run on clean baseline still exits 0.
- Adding a phantom 'PhantomDivergence' entry to baseline.cross_spec_
  duplicates causes the validator to exit 1 with a clear "remove
  from baseline" message naming the stale key.