perf(contains-pr12): batch eligible Contains filters by default - #12999

A length filter was the one shape the classifier turned away for a reason
that was not about the data: lengths land in a roaringset bucket, encoded
by the same lexicographically sortable int64 the primitive int family
uses, so a length batch differs from an int batch only in which bucket it
names. Batching it removes a family from the desugared path without
changing what any query answers.

No supported API can send this shape today. ValidateFilters admits only
(not) equal and greater/less than (equal) on a length path, so a len()
Contains filter is refused before it reaches the searcher; the arm becomes
reachable when that whitelist widens. An acceptance case pins the refusal,
so widening it fails there until the shape gets coverage of its own.

The leaf therefore has to carry the length property name rather than the
property's own. newBatchedContainsPair took the property and derived the
name itself, which left every family sharing one rule; it now takes the
name, so each builder states the bucket it reads and the length builder
cannot inherit the wrong one. One int builder serves both, since only that
name differs.

Bucket presence is not enough to accept the filter. The desugared path
refuses a length filter while the class does not index lengths, so a batch
reading the bucket directly would answer where that path fails, and the two
would disagree on the same query against the same data. The classifier
checks the flag, so the filter desugars and reaches the existing refusal.

A property whose own name ends in the length suffix is not a length filter,
and readFromBucket cannot tell the two apart: it decides from the suffix
alone. The classifier applies the same test, so batching does not start
answering a filter every released version refuses. That refusal is itself
wrong — the property's own bucket holds the values asked for — and fixing
the suffix test belongs to the path that owns it.

The differential test grows a length arm: several present lengths, every
present length, a duplicated one, present-with-absent and all-absent,
across all three operators. A document holds one length, so ContainsAll
over two distinct lengths is empty by construction and a duplicated
length is the only non-empty ContainsAll — both compared against the
desugared compound. Pointing the length builder at the value bucket
fails that arm and the unit test. Each row resolves through a helper that
fails unless the batched path ran, so an arm that stops batching fails
instead of comparing the desugared path with itself.

Two fixture classes gain an InvertedIndexConfig. Every class reaching the
searcher through a creation path has one, and the property-length and
null-state guards in readFromBucket dereference it without checking, so a
fixture without one crashes the moment a test filters on a length. No load
path fills it, which the classifier declines for rather than dereferencing.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

queryIDsErr formatted resp.Errors with %v, and that slice holds pointers, so
every failure read "graphql errors: [0x327b3101a080]". A caller asserting on
the message text had nothing to match.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>