refactor(server): attach the SSE keepalive through one constructor

[inureyes]

Summary

Five route handlers each ended with the byte-identical tail Sse::new(stream).keep_alive(keepalive.into_inner()).into_response(). A sixth route that forgot the .keep_alive(..) would have compiled and passed the whole suite. This replaces the five copies with one shared constructor, so forgetting the keepalive stops being expressible, and deletes the TODO that asked for an axum integration harness instead.

Why not the test the TODO asked for

The TODO at streaming_tests.rs named a real risk, but the property splits in two and neither half wants that test. Whether axum emits a comment frame while the stream is idle tests axum's KeepAlive, which upstream already covers. Whether every route attaches the keepalive is the half that can actually regress, and an integration test catches it only for routes someone remembers to write a test for. That is five tests guarding five copy-paste sites: the same duplication, moved into the test module, and still blind to the sixth route.

What changed

• src/server/streaming.rs: new IntoKeepAlive trait and sse_response constructor. sse_response takes the keepalive by value, so a route cannot call it without one.
• src/server/streaming_responses.rs, src/server/streaming_anthropic.rs: a one-line IntoKeepAlive impl each. The three newtypes stay distinct, which is the property refactor(server): consolidate the three duplicated SSE keepalive constants under one proxy-timeout assertion #1105 asked to keep: a route still cannot attach another surface's keepalive, because the value it holds came from the channel constructor it called. The trait is the only thing they share.
• The five route tails become sse_response(stream, keepalive).
• src/server/router_front.rs: its two sites go through the same constructor. They are outside src/server/routes/ and so outside the letter of the criterion, but they were the last hand-assembled Sse::new in the tree, and including them is what makes the one-attach-site property hold tree-wide rather than per-directory.
• src/server/streaming_tests.rs: the TODO is replaced by a note recording that the hazard went away structurally.

Acceptance criteria

Criterion	Evidence
No Sse::new(..).keep_alive(..) remains in src/server/routes/	grep -rn "keep_alive" src/server/routes/ returns nothing
All three keepalive newtypes reach the shared constructor	One IntoKeepAlive impl each; all five routes call sse_response
Deleting .keep_alive(..) from the shared constructor is the only way to break the invariant, and it breaks it for every route at once	`grep -rn "Sse::new(" src/
The TODO at streaming_tests.rs is removed	Replaced with a note on why

Deleting the tails made sse::Sse unused in all five route files and clippy flagged it under -D warnings. That is the compiler confirming no route builds an SSE response on its own any more; the imports were removed.

Test plan

• cargo test --profile test-fast --features cuda --lib server::streaming: 25 passed, 0 failed
• cargo test --profile test-fast --features cuda --lib server::routes: 116 passed, 0 failed
• cargo clippy --profile test-fast --features cuda --lib --tests -- -D warnings
• cargo fmt --all -- --check

Behaviour is unchanged: every route emits the same frames on the same schedule, so there is no CHANGELOG.md entry. The acceptance criterion names --features metal,accelerate; this is a Linux CUDA box and cannot build that feature set, and no feature gate reaches this code. The criterion's server:: selector was run as its two populated halves, server::streaming and server::routes, because an unfiltered server selector matches hundreds of tests and the full lib suite aborts under parallel execution on this CUDA host for unrelated reasons.

Closes #1107

---

Review round

A review pass verified behaviour equivalence against the vendored axum 0.7.9 source: sse_response expands to the byte-identical expression the five tails carried, with no boxing, buffering, or adapter inserted, and the error type narrowed rather than widened (all seven call sites feed Infallible). It also confirmed the three newtypes remain non-interchangeable, since default_for_long_prefill is private on both the Responses and Anthropic types, so each is obtainable only from its own channel constructor.

Two things were fixed as a result.

Stale docs, in the file this PR touched most. The hand-assembly sentence was updated in streaming_responses.rs and streaming_anthropic.rs but four more copies remained in streaming.rs. The worst was on sse_channel itself, telling a contributor adding a sixth route to write Sse::new(stream).keep_alive(keepalive.into_inner()) by hand, which is precisely the duplication this PR removes. All four now point at sse_response.

The last public route to a raw KeepAlive is closed. The three into_inner() accessors are gone and the IntoKeepAlive impls return self.0 directly. Each accessor's only caller was its own trait impl, and two were pub inside pub modules, so they were the remaining way for outside code to obtain a bare KeepAlive and hand-assemble an Sse. Raised as optional in review; taken because this PR's central claim is that hand assembly is unreachable, and a public accessor would have left it resting on convention at the one point it should be structural.

Also folded in: the router_front.rs comments regained the sentence stating that KeepAlive::default() is likewise 15s under axum 0.7.9, which is the evidence for the no-behaviour-change claim at those two sites; the IntoKeepAlive for SseKeepAlive impl moved next to impl SseKeepAlive to match the other two files.

Noted, not done. axum::response::sse::Sse is a public third-party type, so nothing mechanically stops a future route from importing it directly. A clippy.toml with a disallowed-types entry pointing at sse_response would make this PR's claim compiler-checked rather than reviewer-checked. The repo has no clippy.toml, and adding a workspace-wide lint config is a larger decision than this issue asked for; it is recorded as follow-up in the technical report.