docs: make every example compile against the SDK the docs describe

[JosteinGj]

The three Doc tutorials jobs have been failing on every run in recent history. The
harness was right: the examples had drifted from the SDK across roughly a hundred
page/language pairs — 245 distinct compile errors from nine root causes — plus one
real harness bug and one stale plan selection that was hiding a defect.

Job	Before	After
Plans match the docs	7 failed, 340 passed	440 passed
Docs match the SDK's API	7 failed, 215 passed	222 passed
Java and Rust examples compile	100 failed, 58 passed	159 passed

Every claim was checked against source — datahub-java-sdk/datahub-api-model and
dataplatform-rust-sdk at main — rather than inferred from the error text. The local
run was first made to reproduce CI error-for-error (100 failed, 58 passed) so the
before/after means something.

What was wrong

Java. DataWrapper.getItems() returned Collection, so 27 .get(0) reads never
compiled. That envelope is now a List (datahub-platform#141, merged), so those reads stand as
written, and the eight pages that had worked around it with .iterator().next() were brought into
line — 35 sites reading getItems().get(0), the way the Python and Rust tabs index. Timeseries has no static of() and NodeModel's setters are Lombok-generated
and return void, so neither the factory nor the 19 fluent chains after it existed.
ResourceForm and ResourceSearch are gone — resources are created from Resource, and
SearchBody<F>'s own javadoc names ResourceSearch as one of the four classes it folded
together. ResourceNetwork.nodes() is a Set<NodeModel>, so 16 traversals mapped
Resource::getExternalId over a stream that never holds one.

Rust. Event::new takes all three required fields, and the struct's r#type and
event_time are plain values rather than Options — the field's doc comment says why —
so 27 pages passed the external id alone and then assigned Some(...) to both.
Node::external_id is a method. FileUpload::new_with_destination_path answers
io::Result. Datapoint.value is Option<f64> and .timestamp a DateTime<Utc>, so
five pages parsed strings that were never strings. BasicEventFilter is now EventFilter
(the criteria) and EventFilterForm (the request body). The tutorial's step blocks showed
none of the use lines they needed, so a reader following the steps could not compile them.

Python. events.filter is keyword-only, and there is no BasicEventFilter; seven pages
used the Rust shape. docs/reference/events.md already described the Python call correctly.

Two that are not doc fixes

A harness bug (test(compile): a reader's value cannot fail the call it is passed to).
javac resolves neither an overload nor a target type through an argument it could not
resolve, so an unresolved reader-supplied name is reported twice: as the name, and again as
a type error on the enclosing call — a different line, which is why the existing per-line
suppression never caught it. Five pages were failing on calls that compile. The stubbed pass
now retracts as well as adds, and narrowly: typing errors only, only in units that have
placeholders, only where the stubbed pass is clean at that line. Lookups are untouched, so a
missing method or type still fails however the reader's values are typed. The new
overload_with_reader_argument control fails without the change.

A stale plan selection was hiding a bug. reference__timeseries's exclude = [13] was
written for a listing of IngestResult's signatures, which is now block 16 — so it had been
skipping the latest-datapoint example and compiling the prose instead. With the selection
corrected, that example turns out to name a DatapointDTO that does not exist: latest
answers DataWrapper<DataCollection<DatapointString>>.

docs/reference/labels.md, policies.md and tenant.md had runnable Java and no plan.
Each is a reference page of separate examples, one of which deletes what an earlier one
creates, so each gets independent = true — which also clears their two "variable already
defined" compile errors.

Also in this branch

• Rust reads the same way. get_items() lends the vector, so taking the first result in one
  expression meant get_items_mut().remove(0) — a mutable borrow and a destructive take, in order
  to read. The 24 sites now bind the response and index it.
• units.by_external_ids → by_external_id on three pages, its prose and the coverage table:
  the sync Python client was renamed (dataplatform-rust-sdk 96fe3ea) and no longer differs from the
  async and Rust clients.
• timeseries().list(100, 43L) resolves the data set id through datasets().byIds rather than
  naming a bare number.
• The tailings page keeps its if let Some(..). It had grown an unwrap_or("") and an
  is_empty() test purely to satisfy the compile tier; the tier now drops E0282 on a page that
  leaves a function to the reader, which is information rustc cannot have. A control fails without
  that change.

Verified

Run the way CI runs it, all zero-fail: the structure tier under system Python with no SDK
(440 passed), test_compile.py and test_api_surface.py against SDK main and the
platform's default branch, and test_harness.py. Whole suite: 741 passed, 78 skipped — the
skips are the live tutorials, which need a stack.

Left alone deliberately

• value_type = "numeric" on ~6 pages. The type error is fixed; whether the server accepts
  "numeric" rather than the float32/BIGINT used elsewhere is a question for the live tier.
• timeseries().list(100, 43L) — the Java client has no external-id overload, so the example
  passes a numeric id with a comment saying so. Whether the page should show resolving the id
  first is a separate call.
• The tutorials job now reaches its gate for the first time in this history and skips:
  no DOCTEST_BASE_URL is configured, so there is no stack to run against. Nothing here has
  been proven against a live backend, and the two items above are the candidates for red when
  one is configured. That is a different verdict from this one.

🤖 Generated with Claude Code