Move generic expansion out of the per-keystroke path (S7)

[Aaronontheweb]

Stack, bottom to top: S1 #8525 -> S2 #8526 -> S3 #8527 -> object elements #8528 -> S4 #8530 -> S5 #8532 -> S6 #8533 -> F1 #8534 -> F2 #8536 -> F3 #8537 -> S7 #8538. Each PR is one commit on top of the one below it. This PR is S7, the top of the stack.

What changes

Nothing in output or diagnostics. The ManifestPrefix expansion no longer runs inside the serializer's per-keystroke step. F2 and F3 both listed that as their known limit.

• The per-keystroke step records a small spec. For a prefix registration it reads the attribute: the target definition, each type argument classified as fixed, explicit closed set, or discovered closed set, the prefix, any explicit manifest, and AllowEmpty. No walk, no type construction, no schema extraction. An invalid target or argument still reports AKKASG040 with the same text and location.
• A once-per-edit expansion step. It takes the compilation, the prefix specs, and the facts stage's closed sets. It builds every combination, constructs each type, extracts its schema, applies the explicit-manifest override and the derived-manifest rule exactly as F2 did, and produces a table keyed by serializer. Resolve and every report output read the merged result. Nothing downstream changed.
• Order preserved. A first version appended the expansion entries and changed the order of dispatch arms against the golden files. Each spec now carries the index of the attribute it came from, and the merge splices entries back into place, so the output is byte-identical.
• Three new caching scenarios over a fixture with a prefix registration. An unrelated edit regenerates nothing. A whitespace edit inside the generic definition regenerates nothing. A new marked implementor of the expanded protocol regenerates only the owning serializer.

What this buys, and what it does not

The walk is gone from the per-keystroke step, and everything downstream stays cached. But the new step still constructs and extracts every expansion on each keystroke before finding its table unchanged, because it needs the compilation and cannot skip. On a serializer that expands over 100 members, that cost remains. The benchmark's new prefix variant makes it visible for the first time.

Removing it needs expansion by substitution over the cached schemas: the definition's schema and each member's schema are already plain values, so a construction's schema can be derived with no symbol at all. That is proposed as S8 on the design page, and it is the step that would make a prefix registration cost nothing per keystroke.

Cost

The benchmark gains a PrefixVariant switch: one serializer registers Envelope<TProtocol> with ManifestPrefix over its 100 protocol messages. Base is the F3 tip. Allocations are the reliable column; means at short-run resolution on a shared machine carry error bars wider than these differences.

Row	Base (F3)	This PR
Fresh driver, plain corpus	19.07 MB	19.08 MB
One field renamed, plain	9.67 MB	9.64 MB
Comment edit, plain	3.61 MB	3.65 MB
Unrelated file edited, plain	3.59 MB	3.59 MB
Fresh driver, prefix variant	23.47 MB	22.96 MB
One field renamed, prefix variant	14.00 MB	13.53 MB
Comment edit, prefix variant	5.43 MB	4.95 MB
Unrelated file edited, prefix variant	5.37 MB	4.93 MB

The plain corpus is flat. The prefix variant's no-change rows drop about 8 percent, which is the walk. The 1.3 MB gap that remains between the two corpora on a no-change edit is the construction cost described above.

How it was checked

372 tests pass (369 plus 3 scenarios), three runs in a row. Golden, wire, and both model snapshots byte-identical. Generator and Akka.Remote build with warnings as errors. The design record gains an addendum under Decision 18 saying where the walk went. Local runs used NuGetAudit=false during the restore break that #8535 fixed; nothing of that is committed.