fix(sse): serve the stored 429 instead of re-sending refused requests

[maxmad64bis]

⚠️ base-red inherited: #14547

Summary

A 429 that names a real rate limit used to be served again on the next leg of the same wave: each leg re-offered the same refused route, so one refusal could surface as several identical 429s. The wave now serves the stored 429 directly instead of re-sending a request the provider already refused. Scope: the attempt loop only; the opt-in park-and-replay path (OPENCODE_PARK_AND_RESUME, default off) runs outside the loop and is out of scope. When every route is refused, the stored 429 is served as-is (no extra call).

Why this has no flag, unlike the early stop you gated in #13657: that rework keeps every opencode 429 moving to the next account (#9611), and so does this. The request still tries every other route. What changes is that it no longer re-sends through a proxy route that just returned 429 for this same request: the provider answers per egress address (#9611), so another account behind that route gets the same 429. When nothing else is left, the client gets that real upstream 429 untouched, the same contract as your #13657 rework. The only calls removed are ones that couldn't succeed, which also spares the provider a burst of identical refused requests.

Related Issues

• Related to fix(sse): pause failover dispatch after repeated transient upstream failures #13615
• Related to feat(sse): retry transient 5xx backend errors with jitter (global-fallback call site) (#12695) #13143
• Related to fix(resilience): trip provider cooldown on credit-exhausted signals beyond 429 #13041

Validation

• Change type: routing
• Focused tests and category gates from the golden path
• npm run lint — ESLint on the touched files is clean (exit 0 with repo suppressions); the full run is red on the base (🔴 Release branch not green: release/v3.8.51 #14547 — cliproxyAccountHealth.ts TS2322, file outside this diff, identical on base)
• Reconciled with the current active release base (18bbb101, focused checks rerun afterward: file-size green, complexity 0, eslint 0, unit 3/3 + neighbors 22/22)
• Production-code changes include a new or updated automated test in this PR

Tests Added Or Updated

• tests/unit/opencode-429-proxy-dedup.test.ts (new, 3 tests): refused route served once instead of re-sent (2 accounts/1 route → 1 call), all-refused fail-open (stored 429 served, no extra call), same-host distinct-port untouched (2 calls, 200 served). Park + early-stop flags unset during tests, restored after.
• Neighbors rerun, no regression: opencode-429-park-resume.test.ts 8/8, opencode-transient-rotation.test.ts 11/11, opencode-accounts-per-request.test.ts 5/5.

Coverage Notes

• Touches open-sse/executors/opencode.ts only (+ test): the 3 new cases cover the stored-429 path (write, read, guard). No coverage moved down in any touched file.

Reviewer Notes

• typecheck:core single error src/lib/services/cliproxyAccountHealth.ts(157,5) is base-red (file identical to base, out of scope); zero errors on the touched files. File-size gate is green locally (opencode.ts at the frozen line count, net zero vs base, re-verified after the base advance to 18bbb101).
• API Route Typecheck red is inherited, not from this PR: the 5 ✗ lines reproduce on the bare base (git diff $mb empty on all three files), same family as test(sse): lock per-key ceiling 429 out of neighbor cooldown scope #14581/feat(sse): distinguish content-stall from transport failure in stream error codes #14582. Docs Gates red is likewise inherited (DEEP_HEALTH_CHECK_ENABLED env-contract gap, diff $mb empty on docs). Unit shards red is unstable, not from this PR: this PR's own test never appears in any shard failure log (0 hits on all 4 shards), local re-runs are 3/3 + neighbors 8/8 + 11/11 + 5/5 green. Fast Quality Gates mutation drift covers third-party modules only (opencode.ts is not a mutated module, 0 hits locally). Base-red 🔴 Release branch not green: release/v3.8.51 #14547 cited.
• ⚠️ base-red inherited: 🔴 Release branch not green: release/v3.8.51 #14547

[diegosouzapw]

Nice, targeted fix — avoiding a re-send through a proxy route that the provider just refused
with a real rate-limit 429 (instead of burning another call that can't succeed) is a clear
win, and the scope cut (attempt loop only, park-and-replay explicitly excluded) is exactly
right. The dedup test spins up real TCP listeners and drives the executor end-to-end rather
than mocking around it — good coverage. Only note: you also have #14659 open in the same
area (pool re-selection after a per-address 429) — worth a quick cross-check before both land
to make sure they don't step on each other.