fix(sse): gate structural chat admission shedding on real heap pressure

[diegosouzapw]

Closes #10183, Closes #10268

Root cause

Both issues share one root cause in src/shared/middleware/chatBodyAdmission.ts. 3.8.49 (#9654/#9940) replaced the 3.8.48 heap-ratio shed (heapUsed/heapLimit >= 0.75) with an unconditional CHAT_MAX_HEAVY_IN_FLIGHT=1 structural lease. A second concurrent "structurally heavy" chat request (messages >= 200, tools >= 64, or estimated tokens >= 32000 — routine for coding-agent fan-out such as Hermes/Cursor/Claude Code) is hard-rejected with a retryable HTTP 503 chat_admission_busy / structure_limit after a bounded queue wait, regardless of actual heap pressure, even on a host with ample free RAM.

• [BUG] Regression 3.8.48 → 3.8.49: chat admission rejects concurrent requests with 503 "Structurally heavy..." even on a healthy heap #10183: reporter confirmed downgrading to 3.8.48 eliminates the error; two concurrent heavy requests on a healthy heap → second is shed.
• [BUG] API call failed (attempt 1/3): InternalServerError [HTTP 503] #10268: Hermes Agent / Cursor fan-out hit the exact same gate; the client reads the fabricated 503 as an upstream capacity error and burns its retries.

Fix

Re-gate heavyweight shedding behind a real heap-pressure probe, layered on top of (not replacing) the #9654 bounded-concurrency / per-connection-lane protection and the #4380 heap-amplification-OOM valve:

• Fast path unchanged: if heavyweight capacity is free, admit immediately.
• When capacity is busy, only enter the bounded-wait/shed path if heapUsed / v8 heap_size_limit >= OMNIROUTE_CHAT_ADMISSION_HEAP_SHED_RATIO (default 0.75, restoring 3.8.48 semantics). A healthy heap admits the second heavy request immediately via a no-op lease instead of parking or shedding it.
• The probe is injectable (admitChatStructure({ heapPressureCheck })) for deterministic tests; defaults to live process.memoryUsage() / v8.getHeapStatistics().

Regression tests (TDD, Hard Rule #18)

• tests/unit/bug-10183-admission-heavy-healthy-heap.test.ts (new, permanent) — RED→GREEN: a 2nd heavy request on a healthy heap is now admitted; a genuinely pressured heap still sheds it (preserves the [BUG] Large /v1/chat/completions request crashes server with Node heap OOM #4380 valve).
• tests/unit/probe-10268-structural-503.test.ts (promoted to permanent) — the exact reported 503 chat_admission_busy shape is still produced under real heap pressure, and the same Hermes/Cursor-style fan-out is admitted on a healthy heap.
• tests/unit/chat-body-admission.test.ts, tests/unit/chat-body-admission-queue.test.ts, tests/unit/per-connection-admission-9654.test.ts updated to inject heapPressureCheck: () => true where they exercise the busy/shed path, so existing feat(backend): Per-connection virtual admission lanes for agentic workloads #9654/[BUG] Large /v1/chat/completions request crashes server with Node heap OOM #4380 coverage is preserved rather than weakened.

Fail→pass evidence (focused suite, 68/68 passing):

✔ bug-10183: second concurrent heavy request admitted on a healthy heap
✔ bug-10183: a genuinely pressured heap still sheds the 2nd heavy request
✔ #10268: 2nd structurally-heavy agent request is rejected 503 (chat_admission_busy) under real heap pressure
✔ #10268: 2nd structurally-heavy agent request is admitted on a healthy heap (the fix)

Gates run

• npm run typecheck:core — clean
• npx eslint --suppressions-location config/quality/eslint-suppressions.json <changed files> — clean
• node scripts/check/check-file-size.mjs — OK, no new frozen-file growth
• node scripts/check/check-test-discovery.mjs — OK, new test files discovered
• Focused admission suite (chat-body-admission*, per-connection-admission-9654, bug-10183-*, probe-10268-*) — 68/68 passing
• Husky pre-commit (lint-staged, docs-sync, any-budget, tracked-artifacts) — passed
• npm run test:unit (full suite) — run under heavy shared-devbox contention from a concurrent 13-way session fan-out; no admission-related failures observed through the run, sole failure seen was a pre-existing unrelated POST /v1/search proxy timeout consistent with known load-induced flakiness (not touching this file). CI will confirm the full suite green.

⚠️ base-red inherited: #9985 — ESLint errors (2) from #10250 (unrelated i18n PT-PT translation regression, not touched by this PR).