[Router] Pin the backpressure disposition at the dispatch sites

[Kangyan-Zhou]

Motivation

Follow-up to #39464 (merged as 4e0b56c8), which taught the circuit breaker
to treat an upstream 503/429 as backpressure rather than a fault.

That PR's end-to-end tests asserted only that a 503 leaves the breaker Closed.
record_success satisfies that identically — so the dispatch-site wiring was
not actually pinned. Verified by mutation on the merged commit: replacing
BreakerOutcome::Neutral => breaker.record_backpressure() with
record_success() in both dispatch arms passed the entire suite, all 918
tests, undetected.

The property that makes Neutral a third verb rather than an alias for success
is that it does not reset an in-progress failure streak — "a worker
interleaving genuine 5xx faults with 503s still trips". That was covered on the
breaker alone (backpressure_in_closed_state_preserves_failure_streak), never
through forward_json_to / forward_streaming_to.

Modifications

Tests only — no behaviour change.

A scripted mock worker. spawn_scripted_worker serves a sequence of
statuses, repeating the last. A fixed-status worker structurally cannot express
the interleaving that separates the two verbs.

Three tests that fail if a dispatch site records backpressure as success.
Each drives fault, fault, backpressure, fault against an explicit
threshold of 3 and asserts the final fault still opens the breaker:

test	arm
engine_503_does_not_reset_an_in_progress_failure_streak	JSON
engine_429_does_not_reset_an_in_progress_failure_streak	JSON — 429 shares the Neutral disposition but only 503 had ever been exercised end to end
engine_503_does_not_reset_an_in_progress_failure_streak_streaming	streaming, which records on a separate code path (up front, skipping the pump hook)

The threshold is pinned via STREAK_THRESHOLD rather than inherited from
CircuitBreakerConfig::default(), because these assert on an exact streak
length — a change to the default must not silently move the trip point.

Two documented-but-untested states.

• backpressure_while_open_does_not_reset_the_breaker — record_backpressure's
  Open arm is justified at length in its docstring and had no coverage. Asserts
  a late 503 neither untrips the breaker nor restarts its cool-down.
• engine_503_recovers_a_half_open_breaker_streaming — the JSON arm had a
  half-open wedge guard; the streaming arm did not, and the probe can just as
  easily land there.

Also fixes the unresolved [`record_success`] intra-doc link in the same
docstring (Self::-qualified, matching its siblings).

Tests

cargo test green — 730 lib, 7 bin, 65 component, 121 proxy.
cargo clippy --all-targets -- -D warnings clean; pre-commit clean.

Mutation-checked in both directions: green as committed, and with
record_backpressure → record_success re-applied at both dispatch sites
exactly the three streak tests fail.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation — N/A, tests only.
• Provide accuracy and speed benchmark results — N/A, tests only.
• Follow the SGLang code style guidance.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Wt25p6zfdw6QFLguPRsj82

---

CI States

Latest PR Test (Base): ✅ Run #35379854356
Latest PR Test (Extra): ❌ Run #35379854013
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.