Milestone 3.5.F.1 — End-to-End Learning-Loop Test SuiteCopy Anchor Link

Status: Completed Goal: Track F — Exit Gate
Phase: 3.5 — Extraction & Context Assembly
Estimated effort: Composition complexity comparable to 3.5.D.4 (not the original 3-day sketch)
Track: Track F — Exit Gate
Depends on: All Tracks A–E milestones

---

OutcomeCopy Anchor Link

Prove the composed learning loop against process-managed services on compose-test infra (Postgres/Redis/ClickHouse), not a fictional full app mesh in compose-test.yml:

make e2e-smoke-p3.5 → infra/scripts/e2e_phase35.sh (IBEX_E2E_P35_MANAGE=1)

ScenariosCopy Anchor Link

1. Extraction → injection — terminate with non-empty turn buffer → worker stub extraction → memory row → next chat injects (X-IBEX-Memories-Injected / marker)
2. Cross-tenant smoke — Org A marker absent from Org B search/chat (thin; full matrix is 3.5.F.2)
3. Latency mechanism — 20 concurrent Assemble; print p95/p99 (soft; not CI hard threshold)
4. Proxy overhead mechanism — Assemble-on chat headers + optional ibex metadata timings printed (soft)
5. Degradation (ADR-0071) — L0–L1 require candidates_evaluated > 0 (L1 via skip_cold); L2 hard zeros with skip_hot+skip_cold; L3 via X-IBEX-Context-Fallback=true (no 4-valued header; no was_truncated proto field)
6. MCP/proxy consistency — write_memory → eventually proxy injects same substrate

Harness notesCopy Anchor Link

• Extraction uses OpenAI-compat HTTP stub (phase35_e2e_extraction_stub.py); worker stays IBEX_WORKER_EXTRACTION_PROVIDER=openai
• Uniqueness markers are Presidio-verified word phrases (phase35_e2e_safe_marker.py) — hex/NATO concatenations false-positive as NRP/PERSON and break search/inject proofs
• Sync via bounded eventually() polling of memory/search/headers — not Celery AsyncResult (task_ignore_result=True)
• CI job display name e2e-smoke-p3.5 (workflow id e2e-smoke-p35; YAML job ids cannot contain .)
• Prefork worker asyncio: _run_coro() in session store / engine dispose (Celery + asyncio.run loop mismatch)

Observed latency (local compose-test, soft)Copy Anchor Link

Recorded from a green make e2e-smoke-p3.5 run (mechanism-gated; not CI hard thresholds):

• Scenario 3 Assemble (n=20): mean ≈ 249 ms, p95/p99 ≈ 286 ms
• Scenario 4 proxy overhead headers: p95/p99 ≈ 314 ms; context assembly p95 ≈ 183 ms

Success signalsCopy Anchor Link

• make e2e-smoke-p3.5 exits 0 (exit criterion #9)
• Six scenarios exercised with empty-buffer hazard addressed (memory/marker proof after terminate)
• Degradation assertions match ADR-0071 (not sketch L0–L3 headers); L0/L1 require non-zero candidates
• Additive scripts/Makefile/CI — F.1 harness is new; collateral fixes only where older suites were broken
• Repo guards / CI checks still pass

PrerequisitesCopy Anchor Link

• All Tracks A–E milestones