Derive the StreamOne ETag from the numeric revision for revisioned documents

[uniquelau]

Addresses #5120.

What

StreamOne<T>'s ETag gate only reads the Guid mt_version flavor — which projection-target documents are forbidden from having (numeric revisions are forced by ProjectionDocumentPolicy, by design per #2978/#3057/#3097).

Net effect: the common CQRS read-model shape (a SingleStreamProjection document served by StreamOne<T>) can never emit an ETag, even though its mt_version already holds the source stream's version — the exact value StreamAggregate<T> serves as its ETag.

Commit 1 teaches the ETag path numeric revisions: StreamOneJsonResult widens to carry either flavor, revision mode piggy-backs the same VersionSelectClause/mt_etag_version select (no new SQL machinery; column read robustly at either width per #4614), and WriteSingle formats via the existing ETagHelpers.Format(long) — the If-None-Match/304 branch is shared code, so behavior is identical to the Guid path by construction. For a single-stream target the ETag equals what StreamAggregate serves for the same stream, so the two read styles are cache-compatible (Inline lifecycle; an Async-lifecycle document can lag the stream head).

Commit 2 (separable — say the word and we split it into its own PR referencing #2978): CompileAndValidate now throws an actionable InvalidDocumentException when both Metadata.Version.Enabled and Metadata.Revision.Enabled are set. Today that state surfaces as a raw duplicate-key ArgumentException from projection storage, or invalid DDL with two mt_version columns — DocumentStorageDescriptorBuilder already documents the never-both assumption. The only fluent sequence newly rejected is UseNumericRevisions(true) followed by UseOptimisticConcurrency(true) on one store, which never worked (duplicate-column DDL). We deliberately did not make UseOptimisticConcurrency(true) clear Revision.Enabled: that would route a projection target past the guard into the Guid-only mapping that breaks on the second inline apply (ConcurrencyException — the projection path binds DBNull as the expected version), a silent runtime failure instead of a bootstrap error. We also considered and rejected moving ProjectionDocumentPolicy to post-policies — it would silently override explicit user configuration and reorder ConfigureAggregateMapping.

Known non-goals / scope notes

• The compiled-query overload (StreamOne<TDoc, TOut>) does not participate in ETag/304 handling (pre-existing; the same VersionSelectClause piggy-back would extend it later if wanted).
• Revision-derived ETags are deterministic: a projection-logic change + rebuild that doesn't advance the stream won't invalidate cached responses (ProjectionVersion bumps don't reset the revision) — matching shipped StreamAggregate semantics; documented.
• Multi-stream/ILongVersioned targets emit their own monotonic per-document revision — a valid ETag, but not a stream version; documented.
• Malformed If-None-Match tokens behave exactly as on the Guid path (never match → 200) — inherited, unchanged.
• EventProjection outputs are not forced into revisioning and emit no ETag unless they opt in; documented.

Tests

4 new tests beside the existing ETag pins in Marten.AspNetCore.Testing (projection-target emits "1", honors 304, advances to "2" with stale-tag 200; plain IRevisioned document emits its revision), 2 failing-first guard tests beside Bug_2978_*, and the existing no-ETag pin extended to state the full negative condition (its subject has neither flavor).

Suites (net10.0 locally; net9 deferred to CI — no local runtime): Marten.AspNetCore.Testing 101/101 · DocumentDbTests 1089/0 · EventSourcingTests 1604/0 · DaemonTests 260/0 · LinqTests 1421/0 · CoreTests 515 passed with 4 failures, all triaged as not-ours: Bug_4185_codegen_conflict… and rolling_range_partitioning… reproduce identically on clean master (pre-existing), and Bug_4187_ancillary… + divergent_application_assembly… pass in isolation on this branch (full-suite interference).

🤖 Generated with Claude Code