feat(api,ui): showcase workspace — preserve, restore, and replay showcase runs

[w7-mgfcode]

Summary

The /showcase page is a fixed demo runner: every run creates model runs, scenario plans, registry aliases, and on-disk artifacts, but nothing records what was created — objects accumulate unlabeled in the DB and the only run memory is a localStorage FIFO of 5 entries (frontend/src/pages/showcase.tsx:166). The UI exposes only 3 of the 8 seeder presets even though the backend already accepts the full ScenarioPreset enum (app/features/demo/schemas.py:57). There is no way to name a showcase session, find its objects afterwards, or re-run an earlier configuration — the demo cleanup step intentionally deletes nothing (app/features/demo/pipeline.py step_cleanup), so preservation exists de-facto but with zero semantics or discoverability.

Approach

Additive, single-host, no new services or runtime dependencies. The demo slice gains its first persistence: a showcase_workspace table (new app/features/demo/models.py, batch-slice precedent) shipped via Alembic, storing name, seed, preset, scope, and the created run_ids / scenario_ids / batch_id / aliases / artifact paths as the pipeline records them. DemoRunRequest (HTTP + WS start frame) gains additive Optional fields only — preservation policy and workspace name — so legacy clients keep working byte-compatibly. Showcase-saved scenario plans reuse the existing GIN-indexed scenario_plan.tags column (["showcase", "workspace:<name>", "source:showcase"]). The frontend ScenarioPicker exposes all 8 presets as guided cards (backend change ~zero) and adds workspace save/restore UI. The linear streaming pipeline, the WS event taxonomy, and the seeder HTTP schema do NOT change shape; restore/replay re-submits a recorded configuration through the existing run path.

Decomposition

• E1 — Foundation (blocks E2–E4): workspace persistence backbone — feat(api): showcase-workspace E1 — workspace persistence backbone (table + migration + DemoRunRequest fields + pipeline recording) #390
• E2 — Parallel: full preset exposure — feat(ui): showcase-workspace E2 — full preset exposure (all 8 ScenarioPreset cards + expected-skip semantics) #391
• E3 — Parallel: workspace-tagged scenario plans — feat(api,ui): showcase-workspace E3 — workspace-tagged scenario plans (pipeline tags + planner library filter) #392
• E4 — Parallel: workspace restore/replay — feat(api,ui): showcase-workspace E4 — workspace restore/replay (list/load workspaces + replay through the run path) #393
• E5 — Release gate (closes after Foundation + all Parallel): e2e dogfood across presets + docs (API_CONTRACTS, RUNBOOKS, DOMAIN_MODEL) + regression coverage — feat(repo): showcase-workspace E5 — release gate (8-preset dogfood + doc sweep + umbrella close-out) #401

Out of scope (explicit)

• Advanced seed config UI (25+ DimensionConfig/FactsConfig knobs) — reason: requires a new seeder HTTP schema surface; its own epic, while seed + date window are already exposed today.
• Interactive per-phase configuration (mid-run panels) — reason: re-architects the linear streaming pipeline and contradicts the request/response design.
• Artifact export bundle under artifacts/showcase/<workspace>/ — reason: no codebase grounding; disk-layout design unsettled.
• Ops/approval-history phase enrichment — reason: /ops/* endpoints already cover most of it; the remaining need is unclarified.
• Manual winner selection mid-pipeline — reason: no mid-run interactive gate infrastructure exists outside the agent HITL step.

Success criteria

• A showcase run started with a workspace name creates exactly one showcase_workspace row recording seed, preset, scope, created run_ids, scenario_ids, batch_id, aliases, and artifact paths
• All 8 ScenarioPreset values are selectable on /showcase and each runs to a green pipeline (or a documented expected-skip)
• Showcase-saved scenario plans carry ["showcase", "workspace:<name>", "source:showcase"] and are retrievable via GET /scenarios?tags=workspace:<name>
• A prior workspace can be restored: its config reloads into the UI and a replay with the same seed/preset completes without 409/500
• Legacy clients are unaffected: a start frame without the new fields behaves byte-identically to today (demo_minimal, ephemeral)
• Alembic migration applies and downgrades cleanly on a fresh DB; all five CI gates green

Risks

Risk	Mitigation
Replay collides with registry duplicate detection	#146/#324 verified fixed; add a replay regression test that runs the same config twice
SPARSE / short-window presets fail mid-pipeline (NaN WAPE, 75-day backfill gate)	expected-skip semantics per preset card + runbook entry; pipeline already skips NaN folds
WS start-frame change breaks legacy clients	additive Optional fields only; contract test for the legacy frame
demo slice's first DB table triggers an alembic cold-boot import cycle	follow the batch-slice models.py precedent + lazy-import rules in docs/_base/ARCHITECTURE.md
localStorage run history diverges from server workspaces	server is source of truth in workspace mode; localStorage stays for ephemeral runs only

Tracking

• Source of truth: docs/flow-pack-methodology.md + working state .flow/state.md (brainstorm Round 4)
• Milestone: showcase-workspace
• One-pass confidence: 8/10 (backend already accepts all 8 presets and replay blockers are verifiably fixed; restore-vs-replay semantics need design in the E4 PRP)