StreamOne<T> ETag: inline mt_version (eliminate 2nd round trip) + reverse where T:notnull + ETag test hardening (pre-9.18, follow-up to #5015)

[jeremydmiller]

Follow-up to #5015 (ETag / If-None-Match support). To be done before the 9.18 release.

Problem

StreamOne<T> currently computes its ETag by, after the document JSON is already fetched and buffered:

1. deserializing the buffered JSON back into a T (session.Serializer.FromJson<T>(stream)), then
2. issuing a second DB round trip — IQuerySession.MetadataForAsync<T>(entity) — purely to read mt_version.

So with EmitETag = true (the default), every StreamOne request does two DB round trips instead of one, and a 304 saves only bandwidth — the document was already fetched. (This is unlike StreamAggregate, where FetchStreamStateAsync reads the version cheaply before the expensive fold, so a 304 skips real work.)

This second call is also the reason StreamOne<T>/WriteSingle<T> had to tighten to where T : notnull (because MetadataForAsync<T> is constrained notnull).

Goal

Fetch mt_version inline with the document in the original single round trip — same idea as the count(*) OVER() as total_rows stats-column trick that #5014 uses — so:

• StreamOne stays at one DB round trip with EmitETag = true on by default (which the maintainer has OK'd once this lands), and
• the second MetadataForAsync call is removed entirely.

Sketch

• Have the StreamOne streaming path select mt_version alongside the data column (extend the Marten core streaming query used by WriteSingle, analogous to how StatsSelectClause appends count(*) OVER()), and read the version off the same DbDataReader before/while streaming the body — no re-deserialization, no metadata query.
• Once the version no longer comes through MetadataForAsync<T>, the where T : notnull constraint on the public StreamOne<T>/WriteSingle<T> signatures can be dropped cleanly (see the separate constraint-reversal that ships in feat: ETag / If-None-Match (304) support on StreamOne and StreamAggregate (fixes #5010) #5015 itself).
• Guard the mt_version-disabled case: if the doc type has version metadata turned off, skip emitting an ETag rather than emitting a constant "00000000-…" (which would risk false 304s). Confirm what the inline read returns in that config.

Acceptance

• StreamOne with EmitETag = true executes exactly one DB command (assert via command counting / query logging).
• No behavioral change to the emitted ETag or the 304/200 semantics vs feat: ETag / If-None-Match (304) support on StreamOne and StreamAggregate (fixes #5010) #5015's tests.
• Version-metadata-disabled doc types emit no ETag (documented).

---

Also in scope for this follow-up (folded in from #5015 review)

Reverse the breaking constraint tightening. #5015 shipped where T : notnull on the public StreamOne<T> and WriteSingle<T> signatures — a source-compatibility break — solely because IQuerySession.MetadataForAsync<T> is constrained where T : notnull. Once this issue removes the MetadataForAsync call in favor of the inline mt_version read, that constraint is no longer needed and must be dropped from both StreamOne<T> and WriteSingle<T> to restore the pre-9.18 signature. (Loosening a constraint is non-breaking, and this ships before the 9.18 release, so the tightened form never reaches a release.)

Add the ETag test cases identified in review (additive; can land with this work or separately, but before release):

• ETagHelpers unit coverage for the logic branches: * wildcard, W/ weak validator stripping, and multi-value comma-separated If-None-Match lists.
• The 404 path (missing document for StreamOne, missing stream for StreamAggregate).
• The version-metadata-disabled / no-version-column case → assert no ETag header is emitted (guards against a constant "00000000-…" ETag causing false 304s).