fix(repo): reliability E6 — release gate (showcase_rich dogfood + umbrella close-out)

[w7-mgfcode]

Sub-issue of #380 (umbrella: fix(repo): platform reliability hardening — agents, config, ui, forecast). Release gate — closes only after Foundation (#334) + all Parallel epics (#335, #332, #268, #237) close. All five are closed as of 2026-06-12 — this epic is unblocked.

Purpose

Full-gate verification that the five reliability fixes hold as one system on dev, not just as isolated epic PRs: a green end-to-end showcase_rich dogfood run, a per-epic spot check, and the umbrella close-out. This is the last open success criterion on #380 ("showcase_rich demo pipeline runs green end-to-end after E6").

Sub-tasks

• Fresh stack: docker compose down -v && docker compose up -d && uv run alembic upgrade head
• Run /showcase with scenario=showcase_rich + Re-seed first — all 24 steps / 10 phases green (PRP-41 layout; provider-dependent steps may ⏭️ skip per RUNBOOKS, pipeline still green)
• Spot-check each epic on dev:
  • E1 fix(config): reject doubled provider prefixes in agent model ids #334 — PATCH /config/ai with a doubled provider prefix (google-gla:google-gla:…) → 422
  • E2 fix(agents): surface fallback model failures with actionable details #335 — exhausted fallback chain returns 502 AGENT_FALLBACK_EXHAUSTED with classified failures[] (REST) / error_type="fallback_exhausted" (WS)
  • E3 fix(ui): avoid crypto.randomUUID crash on LAN HTTP showcase #332 — /showcase over plain-HTTP LAN origin completes without white-screen
  • E4 refactor(forecast,registry): move ModelFamily to shared/core to break slice cycle #268 — ModelFamily resolves from app/shared/model_taxonomy.py; no lazy-import NOTE references the old registry↔forecasting cycle; alembic cold-boots clean
  • E5 fix(forecast): model_exogenous re-forecast is inert to price assumptions (0.0 delta) #237 — fresh seed → train regression on a real grain → price-cut simulate → method=model_exogenous, units_delta != 0.0
• All five validation gates green on dev (ruff, format, mypy, pyright, pytest)
• Record the evidence on this issue, tick the last success criterion on fix(repo): platform reliability hardening — agents, config, ui, forecast #380, close fix(repo): platform reliability hardening — agents, config, ui, forecast #380

Exit criteria

• showcase_rich dogfood green end-to-end (evidence: step summary or screenshot)
• All five per-epic spot checks pass
• Umbrella fix(repo): platform reliability hardening — agents, config, ui, forecast #380 closed with a close-out comment linking the evidence

[w7-mgfcode]

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
🔍 E6 Release Gate — evidence (dev @ aca758f)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

📋 Fresh stack (E4 cold-boot proof)
✅ fresh empty DB → uv run alembic upgrade head exit 0 (full migration chain, no import cycle)
✅ scripts/check_db.py — connectivity + pgvector OK

📋 showcase_rich dogfood — LAN HTTP origin http://10.0.0.226:5173/showcase (E3 surface)
✅ page rendered: isSecureContext=false, typeof crypto.randomUUID === "undefined", 0 console / page errors
✅ 24/24 steps, zero ❌ — pipeline_complete: winner prophet_like, WAPE 15.1108, 31 s wall-clock
✅ knowledge phase live: rag_index_subset 5/5 files / 43 chunks, rag_retrieve_probe top-hit score 0.642 (Ollama qwen3-embedding:4b)
⏭️ verify — V2 winner artifact-root mismatch (RUNBOOKS item 9, sanctioned)
⏭️ agent_hitl_flow — agent ran (3356 tokens, gemma4-agent) but didn't call save_scenario (item 25, sanctioned)
📸 screenshot /tmp/e6-showcase-final.png · WS step log /tmp/e6-ws-events.jsonl (local gate artefacts)

🧪 Per-epic spot checks
✅ E1 #334 — live PATCH /config/ai doubled prefix → 422 problem+json with the "Did you mean 'google-gla:gemini-2.0-flash'?" hint; 9/9 doubled/mixed-prefix tests green (schema + route + Settings-boot paths)
✅ E2 #335 — test_chat_fallback_exhausted_returns_502_problem_json PASSED; live probe: 502 code=AGENT_FALLBACK_EXHAUSTED, failures[] classified reason=model_not_found (404), no secret values; config restored + verified afterwards
✅ E3 #332 — uuid-utils vitest 3/3 incl. the LAN-HTTP shape; full run over non-secure origin without white-screen (above)
✅ E4 #268 — app.shared.model_taxonomy resolves TREE/ADDITIVE/BASELINE; zero lazy-import NOTEs reference the old registry↔forecasting cycle (only the sanctioned forecasting↔jobs pair at app/features/forecasting/service.py:1050); cold-boot proof above
✅ E5 #237 — test_seeded_train_simulate_price_cut_moves_demand PASSED (engineered elastic grain → method=model_exogenous, units_delta != 0); live: seeded sales_daily.unit_price tracks price_history windows (coupling active, e.g. grain (10,23) = 9.39/9.98). Note: live sims on organic showcase grains return Δ0.0 — lag-dominated learning per the recorded E5 verdict, not the #237 wiring bug

📋 Validation gates (dev @ aca758f)
✅ ruff check · ✅ ruff format --check (363 files) · ✅ mypy app/ (334 files, 0 issues) · ✅ pyright app/ (0 errors) · ✅ pytest -m "not integration" — 1938 passed

⚠️ Rig note: the first dogfood attempt ❌'d at rag_index_subset (embedding ReadTimeout @60 s) — gate-rig fault, not a dev regression: Ollama was started without docker-compose.gpu.yml, so the cold CPU-only qwen3-embedding:4b load exceeded the timeout. With GPU passthrough the warm embed is 2.4 s and the re-run went green.

────────────────────────────────────────────
✅ Result: GATE GREEN — all exit criteria met
────────────────────────────────────────────

PRP file committed via PR #388 (docs/reliability-e6-prp, the E1–E5 precedent).

[w7-mgfcode]

✅ Gate complete — evidence above (#387 (comment)). Umbrella #380 ticked and closed; PRP file lands via PR #388.