release: promote dev to main for v0.2.20#326
Merged
Merged
Conversation
chore(repo): back-merge main → dev (v0.2.19)
Lands the docs-only Forecast Intelligence roadmap — 4 INITIAL docs + 3 PRPs, no production code. Dependency-chained execution: PRP-35 first, PRP-36 + PRP-37 follow. Tracked by epic issue #295. - INITIAL roadmap (A/B/C + index) - PRP-35 Feature Frame V2 — V1 frozen, V2 ships as sibling builders, dispatch at service layer only, load-bearing leakage spec - PRP-36 Model Zoo + Backtesting — new baselines, per-horizon-bucket metrics, comparable-runs with feature_frame_version key - PRP-37 Interactive UI — partial-execution gates, shadcn@4.7.0 pin, per-component @radix-ui/react-X imports
docs(prp): add forecast intelligence planning docs
Bundles three carryover concerns from prior local demo work into one PR.
* fix(data) — PriceHistoryGenerator could emit a row with valid_to <
valid_from when a change roll fired on the window's first day. That
violates ck_price_history_valid_dates and crashed the seeder during
ingest. The fix skips the degenerate row.
* feat(data) — three new local-host scripts that drive the public API
to enrich the demo DB without raw SQL writes:
- seed_phase2_only: re-runs Phase 2 generators (replenishment,
exogenous, returns, lifecycle) against existing dimensions
- seed_historical_activity: submits varied train/predict/backtest
jobs across 2024-Q4 -> 2026-Q1 cutoffs through /jobs
- seed_registry_from_jobs: walks completed train jobs, runs the
canonical pending -> running -> success transition + alias stamps
* chore(repo) — uv.lock refreshes forecastlabai 0.2.18 -> 0.2.19
to match the release-please-merged version bump.
Excluded intentionally: alembic/a2b3c4d5e6f7 + rag/models.py — the
migration is self-marked "local-only demo" (truncates document_chunk,
drops HNSW index, hardcodes 2560 for qwen3) and would wipe any non-
qwen3 user's RAG corpus on upgrade. Stays uncommitted locally.
Three corrections to register_one and fetch_completed_train_jobs:
* pagination — `page * len(jobs) >= total` stops too early when the
last page is partial. Switch to accumulated-count + short-page
detection (exit when len(jobs) < page_size or len(out) >= total).
* model_path validation — empty / directory paths slipped through
because Path("") resolves to cwd and Path.exists() returns True for
directories. Require non-empty path and Path.is_file() for both the
raw and cwd-relative candidates.
* duplicate detection — `r.status_code >= 400` blanket-swallowed
registry downtime and validation errors as idempotent skips. Narrow
the skip to HTTP 409 (the actual DuplicateRunError code per
registry/routes.py:113) and raise RuntimeError on other 4xx / 5xx
with the response body for diagnostics.
Python 3.12-only `def chunked[U](...)` syntax in seed_phase2_only.py
is intentional — `pyproject.toml:6` already pins `requires-python =
">=3.12"`.
…ools feat(data,repo): local demo tooling + seeder price-history fix
Lands V2 feature-frame contract as additive, opt-in surface alongside frozen V1. Training + scenarios + shared builders complete; backtesting V2 dispatch deferred to follow-up tracked in #299. V1 callers unchanged. - Shared layer: V2 manifest (38 default / 53 max columns), sidecars, row builders - Training: TrainRequest gains feature_frame_version + feature_groups (opt-in) - Scenarios: build_future_frame dispatches V1/V2 via bundle metadata - 3 LOAD-BEARING leakage specs land alongside the V1 spec - No Alembic migration (V2 reads existing tables, writes nothing) - V1 bundles load/predict/scenario-simulate/backtest unchanged
feat(forecast): add feature frame v2
Contract Refresh gate (PRP-36 Task 1) executed against dev @ f2bf7c8. Patches PRP-36 to match what PRP-35 actually shipped vs what it assumed. - Add PRPs/ai_docs/prp-35-final-contract-snapshot.md (one-off, authoritative) - Item 7: backtesting V2 dispatch deferred to #299 (not in PRP-35 final scope) - Task 8: re-scoped to V1 fold-path bucket metrics only; V2 lands with #299 - Item 10: DEFAULT_V2_GROUPS order corrected (CALENDAR at position 2) No implementation code touched. PRP-36 execution remains gated until this lands.
docs(prp): refresh PRP-36 after Feature Frame V2
PRP-36 (Forecast Intelligence — Slice B). Promote the model layer from "a regression model + three baselines" to a disciplined model zoo with fair, leakage-safe comparison on top of the PRP-35 Feature Frame V2 contract. Models (under model_factory + _MODEL_FAMILY_MAP): - weighted_moving_average — linear or exponential weighting (always-on) - seasonal_average — average of last N seasonal cycles, optional trim (always-on) - trend_regression_baseline — Ridge over elapsed-day + dow/month one-hots (always-on) - random_forest — sklearn RandomForestRegressor, n_jobs=1, gated by forecast_enable_random_forest Backtesting metrics (additive — V1 fold path only): - aggregated_metrics gains rmse alongside mae/smape/wape/bias - FoldResult.horizon_bucket_metrics — per-bucket dict keyed by h_1_7 / h_8_14 / h_15_28 / h_29_plus (empty buckets dropped) - ModelBacktestResult.bucketed_aggregated_metrics — per-bucket means across folds - V2 backtesting dispatch remains DEFERRED to #299 Registry comparable-run rule: - RegistryService._find_duplicate now distinguishes V1 vs V2 (runs with different feature_frame_version are NOT duplicates) - New find_comparable_runs(grain, overlapping window, same V, status==SUCCESS) - RunCreate.runtime_info_extras lets callers pin feature_frame_version + feature_groups - RunResponse.feature_frame_version + feature_groups computed from runtime_info (legacy runs surface None) Ops staleness: - New StaleReason enum value FEATURE_FRAME_VERSION_MISMATCH — a V1 alias with a newer V2 comparable run reports this instead of NEWER_SUCCESS_RUN - AliasHealth and ModelHealthEntry expose alias_feature_frame_version + comparable_run_feature_frame_version Explainability: - New explainers: WeightedMovingAverageExplainer, SeasonalAverageExplainer, TrendRegressionBaselineExplainer - Factory + service plumb weight_strategy / decay / lookback_cycles / trim_outliers - HGBR keeps raising FeatureImportanceUnavailableError (422 path unchanged) Other: - examples/forecasting/model_zoo_compare.py — read-only diagnostic that backtests every available model on a single grain and prints aggregate + per-bucket WAPE - docs/_base/API_CONTRACTS.md, DOMAIN_MODEL.md, docs/optional-features/05 + 09 updated Validation: - ruff check / format clean - mypy --strict / pyright --strict clean (3 mypy + 8 pyright pre-existing xgboost/lightgbm errors only; CI runs --all-extras) - 1574 non-integration tests pass; load-bearing leakage specs unchanged - alembic check — NO new migration (all new state rides existing JSONB columns) Out of scope (deferred): - V2 backtesting fold dispatch — #299 - PRP-37 UI / dashboard — Slice C - /explain/forecast handler for random_forest — needs bundle reload, separate PRP
CodeRabbit review on PR #303 surfaced one bug-risk + one consistency issue + one missing test + one doc typo + an overall refactor request. All five addressed. 1. BUG-RISK — _run_feature_frame_version returned None for missing JSONB keys while _feature_frame_version_filter treats them as V=1. _alias_staleness compared None != 1 and spuriously surfaced FEATURE_FRAME_VERSION_MISMATCH for a legacy alias against an explicit-V=1 comparable run. Normalized the ops helper to return V=1 for missing keys (matches the registry filter contract). The schema-side RunResponse.feature_frame_version still surfaces None so UIs can distinguish "no V info" from "V=1". 2. REFACTOR — Extracted shared pure helpers in forecasting/models.py: - compute_weighted_average_weights - compute_seasonal_average_for_offset - build_trend_baseline_design_row The forecasters' fit/predict + the three new explainers now call them as the single source of truth. No more two-place drift risk when a default changes. 3. CONSISTENCY — SeasonalAverageExplainer.sample_dispersion now measures the same array the forecast was averaged from (post-trim when trim_outliers is on; raw otherwise). Description updated to match. 4. TESTING — Added test_invalid_min_samples_leaf_raises to round out RandomForestForecaster's constructor-validation branches. 5. TYPO — docs/optional-features/09-…governance.md uses the `stale_reason` field-name form (no hyphen) to match DOMAIN_MODEL.md / API_CONTRACTS.md. Plus: two new ops tests pin the new V=1 normalization contract (`_run_feature_frame_version_rejects_unsupported_value`, `_alias_staleness_legacy_run_treated_as_v1_no_spurious_mismatch`). Validation: ruff / mypy --strict / pyright --strict clean (same 3+8 pre-existing xgboost/lightgbm errors only). 1577 non-integration tests pass (+3 new). Leakage specs unchanged.
…acktesting feat(forecast): add model zoo and backtesting comparison
docs(prp): refresh prp37 after model zoo contracts
PRP-37 — Forecast Intelligence Slice C. Operator-facing surface for the PRP-35 V2 feature contract and the PRP-36 model zoo, backtest buckets, and V-aware ops fields. Backend untouched (per PRP-37); every visible value is read from an existing backend response. Surfaces: - /visualize/forecast: family Tabs + model-type Select + V1/V2 Select + conditional feature-pack toggles + train submission. - /visualize/backtest: per-horizon-bucket table + chart, RMSE tile, baseline-vs-feature-aware comparison table. - /visualize/planner: scenario method badge. - /visualize/batch: 5 sweep presets + multi-model x V1/V2 matrix picker. - /explorer/run-detail: Feature frame panel (V1/V2 + per-group columns + per-column safety chips). - /explorer/run-compare: Champion compatibility badge + Feature frame version row. - /ops: stale-alias V mismatch chip, model-health explainer columns, safer Promote dialog (artifact verify + worse-WAPE ack + V-mismatch ack). Adds 11 components under components/forecast-intelligence/, 1 chart under components/charts/, 2 lib modules under lib/, with colocated vitest tests for every component and helper. api.ts extended with PRP-35/PRP-36 wire types (all Optional, additive). use-runs gains an optional feature_frame_version param (not forwarded to the backend list endpoint; no server-side filter exists). Validation: pnpm tsc --noEmit + pnpm lint + pnpm test --run all clean (202 frontend tests). Backend regression suite (forecasting + backtesting + registry + ops, non-integration) 518 passed.
feat(ui): add interactive forecast intelligence UI (PRP-37) (#305)
PRP-37 introduced `trainFamily` as the train-card form state (useState at L68), shadowing an existing PRP-31 derived `trainFamily` at L120 sourced from `useJobFeatureMetadata`. Babel/Vite reject the file with `Cannot redeclare block-scoped variable 'trainFamily'`; tsc reports TS2451 at both sites. The forecast page does not mount. Rename L120 to `loadedTrainFamily` (distinct semantics: derived from a loaded predict job's training metadata, used by the `ModelFamilyBadge` in the Model details collapsible). The L68 form state keeps the `trainFamily` name.
…licate fix(ui): rename duplicate trainFamily binding in forecast page
…g-lifecycle feat(api,ui): showcase pipeline richer data and v2 foundation (#309)
…ing-artifacts docs(docs): add rich showcase planning artifacts
PRP-39 — extend the showcase_rich demo pipeline with three new decision-
phase steps (champion_compat_compare, stale_alias_trigger,
safer_promote_flow) and a new portfolio phase (batch_preset). The
decision lifecycle now demonstrates V1-vs-V2 champion-compat verdicts,
the stale-alias V-mismatch chip on /ops, and the safer-Promote dialog
gates. The portfolio phase runs the quick_baseline_sweep preset (3
stores x 2 products x 3 baselines = 18 items) via /batch/forecasting.
Backend:
- app/features/demo/pipeline.py — 4 new step functions, PHASE_PORTFOLIO
constant, BATCH_PRESET_QUICK_BASELINE_SWEEP_MODELS module constant,
DemoContext additive fields (compat_compare_result,
stale_alias_run_id, original_demo_alias_run_id, batch_id,
batch_status), step_cleanup extension that restores the
demo-production alias to its pre-swap target (R15).
- app/features/demo/tests/test_pipeline.py — 8 new unit tests (4 step
functions, 2 skip paths, 2 cleanup scenarios) + extended canned
responses for /ops/summary, /batch/forecasting, /registry/runs?...,
/registry/aliases/{name}, /registry/compare/{a}/{b}; lockstep
test_phase_table_showcase_rich expanded to 18 rows.
- tests/test_e2e_demo.py — new test_run_demo_showcase_rich_decision_
portfolio integration test asserting the four new step events fire
and R15 alias restoration completes.
Frontend:
- PHASE_DEFS.ts — appends 3 decision-phase rows + portfolio phase row;
PHASE_ORDER + PHASE_LABEL extend with 'portfolio'.
- showcase.tsx — resolveInspectHref gains 4 new case arms targeting
/explorer/runs/compare, /ops, and /visualize/batch/{batch_id}.
- demo-step-card.tsx — 4 new mini-summary chip-line components.
- demo-step-card.test.tsx (new) — 6 render tests covering chip-lines
and Inspect button behaviour.
- PHASE_DEFS.test.ts + use-demo-pipeline.test.ts — extended to assert
the new 18-step showcase_rich layout.
Docs:
- docs/_base/RUNBOOKS.md — 8 new failure-mode entries under the
/showcase pipeline section covering the 4 new steps (skip / fail
diagnostics, R15 cleanup recovery).
Drift resolutions (per PRPs/ai_docs/prp-39-contract-probe-report.md):
- D1 (compare envelope): champion_compat_compare derives compatible +
comparable_reason client-side; mirrors the frontend
computeCompatibility predicate.
- D2 (quick_baseline_sweep): preset expansion stays in the demo slice
(Option A); no preset_id on BatchSubmitRequest.
- D3 (sync settle): /batch/forecasting normally returns terminal status
on submit; the 90 s poll loop is a safety net.
WebSocket schema additive only — no StepEvent / DemoRunRequest field
changes. Relative-anchor phase insertion (PHASE_PORTFOLIO between
PHASE_DECISION and PHASE_VERIFY) keeps the slice merge-order
independent of PRP-40.
…tfolio-lifecycle feat(api,ui): showcase pipeline decision + portfolio lifecycle (#316)
Adds two new phases to the in-process /showcase demo pipeline on the showcase_rich scenario (PRP-40): planning (2 steps — scenario_simulate_and_save, multi_plan_compare) and knowledge (3 steps — embedding_provider_probe, rag_index_subset, rag_retrieve_probe). Both phases insert BEFORE the verify phase via a relative anchor so PRP-39 (sibling, parallel) rebases cleanly. Backend additive contracts: - IndexProjectDocsRequest.path_prefix: str | None = None on app/features/rag/schemas.py — restricts the docs/ root scan to a sub-path with a path-traversal guard. Default None preserves wholesale-scan behavior. - _parse_artifact_key + _embedding_provider_reachable helpers on app/features/demo/pipeline.py — R16 (scenarios.run_id is the artifact key, not model_run.run_id) and provider-presence-only probe per security-patterns.md. - DemoContext gains scenario_artifact_key / price_cut_scenario_id / holiday_scenario_id / embedding_unreachable fields (None on demo_minimal). Showcase_rich step count: 14 → 19. The knowledge phase SKIPs gracefully when no embedding provider is reachable; pipeline still goes green. Frontend mirrors the lockstep contract: PHASE_DEFS.ts ALL_STEPS gains five new entries + PHASE_ORDER + PHASE_LABEL add planning / knowledge; showcase.tsx resolveInspectHref switch deep-links the new step cards into /visualize/planner (with optional ?scenario_id), /knowledge, and /admin. Five new mini-summary helpers on demo-step-card.tsx render the per-step detail strips. Vertical-slice rule preserved — demo never imports scenarios / rag / config / registry; every call goes through httpx.ASGITransport. Docs: docs/_base/API_CONTRACTS.md notes the new path_prefix field and the two new phases; docs/_base/RUNBOOKS.md adds five new step failure-mode entries to the "Showcase page pipeline fails at step X" section.
…wledge-lifecycle feat(api,ui): PRP-40 showcase planning + knowledge lifecycle (#315)
PRP-41 drift items D1-D7 + fold in issue #311: - D1-D4: KPI strip counters now read the actual emitted step.data keys (completed_items, total_chunks, total_aliases, scenario_id + winner_scenario_id) — never invented names. - D5: GET /ops/model-health takes only ?limit (no grain param). - D6: ops_snapshot derives stale_aliases_count + total_aliases from OpsSummaryResponse.aliases (no flat keys on the response). - D7: replaced stale line citations with symbol refs so the PRP does not re-stale on the next pipeline.py expansion. - Folded issue #311 (phase accordion lock after completion) into scope as polish item 7 + acceptance criterion D10 — load-bearing for the Inspect-Artifacts post-run UX. - Noted issue #312 (phase2 enrichment idempotency) as a dogfood prerequisite (out of PRP-41 scope, must land before the manual checklist runs). Docs only change. No code touched.
docs(docs): refresh initial 41 after prp 39 40 (#313)
A second POST /seeder/phase2-enrichment against an already-enriched scope no longer raises IntegrityError (previously: uq_exogenous_signal_per_store surfaced as HTTP 500, blocking PRP-41's manual showcase_rich dogfood). - exogenous_signal: pg_insert(...).on_conflict_do_nothing() — target-free ON CONFLICT covers both partial unique indexes (global + per-store) - replenishment_event / sales_returns (no natural-key unique constraint): section-level existence check inside the seeded date window; skip the whole insert when rows already exist - product lifecycle UPDATE: already idempotent under a fixed seed; left in place - Phase2EnrichmentResponse gains additive records_skipped: dict[str, int] - Defensive IntegrityError -> ConflictError(409) wrap as the belt-and-braces net (the idempotency guards above should make it unreachable) Adds a non-destructive integration test (app/features/seeder/tests/ test_phase2_idempotency.py) that calls the endpoint twice against the live Postgres and asserts records_skipped > 0 + records_created == 0 on the second pass for all three insert tables.
CI's freshly-migrated DB has only stray data from earlier tests' fixtures — too sparse for the ReplenishmentGenerator or ReturnsGenerator to produce any records. The original assertion required records_skipped > 0 for all three insert tables, which fails when nothing was generated. Relaxed: assert created == 0 on the second call for all three insert tables (the canonical idempotency proof — no new rows). Sanity guard soft-skips when none of the three tables exercised the idempotency path (otherwise the test would be meaningless). The exogenous_signal ON CONFLICT DO NOTHING path — the original IntegrityError surface — is exercised by any DB with at least one Store + one date.
…idempotent fix(data): make phase2 enrichment idempotent (#312)
…t-ops-polish docs(docs): add prp 41 showcase agent ops polish
PRP-41 — fourth and FINAL slice of the /showcase upgrade epic (PRP-38..41). Adds two new pipeline phases on scenario=showcase_rich plus cross-cutting UI polish that closes issue #311. Pipeline (backend / app/features/demo/pipeline.py) - step_agent_hitl_flow: HITL approval round-trip on the experiment agent. Drives POST /agents/sessions + /chat + /approve via ASGITransport; surfaces an intermediate step_complete (status=running, awaiting_approval=true) for the FE to render the Approve button; absorbs 400 "No pending action" when the FE pre-empts; 90 s hard timeout falls back to skip so a hung agent never wedges the run. - step_ops_snapshot: 3 GET calls to /ops/summary + /ops/retraining-candidates + /ops/model-health, derives a 5-key KPI payload (stale_aliases_count, retraining_candidates_count, total_runs, total_aliases, degrading_health_count). warn (never fail) on all-three-failed. - _phase_table() — design Z: unified `agents` phase id for BOTH scenarios; SHOWCASE_RICH swaps step_agent for step_agent_hitl_flow and appends an ops phase carrying ops_snapshot before cleanup. SHOWCASE_RICH = 24 rows / 10 phases; DEMO_MINIMAL = 11 rows (unchanged shape under the new agents phase id). - _Client.yield_event hook + run_pipeline event-sink drain. The orchestrator stamps step_index / total_steps / phase_index / phase_total / phase_name on every drained intermediate event. Frontend (UI) - PHASE_DEFS.ts — design Z restructure: BOTH the legacy `agent` step and the new `agent_hitl_flow` live under the unified `agents` phase id; new DEMO_MINIMAL_ONLY_STEP_NAMES set complements SHOWCASE_RICH_STEP_NAMES so the filter selects the right step per scenario (lockstep test pins 24 tuples / 10 phases). - DemoPhasePanel.tsx — adds onValueChange handler + local useState (closes issue #311 / D10): post-pipeline-complete the operator can finally expand any phase without snapping back to the fallback. - demo-step-card.tsx — HitlFlowSummary chip-line + OpsSnapshotMiniGrid + one-click ApproveButton (only renders when status=running AND awaiting_approval=true). - showcase.tsx — five new chrome additions: - ShowcaseKpiStrip — 5-tile KPI strip above the controls card. - RunHistoryStrip — localStorage FIFO 5 with Replay button. - Stop button (visible mid-run) — closes the WS so the backend's WebSocketDisconnect releases the pipeline lock. - InspectArtifactsPanel — 10 deep-link cards rendered after pipeline_complete. - resolveInspectHref switch extended with agent_hitl_flow → CHAT, ops_snapshot → OPS. - use-demo-pipeline.ts — stop() callback exposed via UseDemoPipelineResult; DemoSummary.v2RunId added (mapped from pipeline_complete event.data.v2_run_id). Docs - docs/user-guide/showcase-walkthrough.md — drops 7 "planned" markers across PRP-38/39/40/41 phases; adds concrete prose for Agents (HITL) + Ops snapshot + the 5 polish items + performance budgets table refresh + screenshot placeholders. - docs/_base/RUNBOOKS.md — 5 new failure-mode entries (23-27): agent_hitl_flow no-key / timeout / no-trigger, ops_snapshot all-failed, Stop button mid-run. Tests - Backend: 9 new tests in test_pipeline.py (HITL: happy / no-key / session-fail / no-tool / 4xx-absorb / timeout + Ops: happy / warn / empty); lockstep test rewrite 23 → 24 tuples; 5 new canned-response fixtures for /ops/* endpoints. - Frontend: 22 new vitest cases across 5 test files (DemoPhasePanel onValueChange, ShowcaseKpiStrip 5-tile derivation, InspectArtifactsPanel 10-card grid, RunHistoryStrip localStorage FIFO, demo-step-card HITL + Approve + Ops mini-grid). - E2E: test_run_demo_showcase_rich_full_epic asserts PRP-41 contract shapes hold when the steps execute; tolerates a pre-existing PRP-39/40 cascade (scenario_simulate_and_save can fail to parse the safer_promote_flow placeholder artifact_uri) documented in RUNBOOKS.md entry 18. Validation - ruff + format clean; mypy + pyright strict (only pre-existing xgboost/lightgbm stub gaps remain — documented in PRP body). - 1635 unit tests pass; 249 frontend tests pass. - Vertical-slice guard empty: zero imports from agents/ops/registry/ scenarios/rag in app/features/demo/. Out of scope (explicit) - No new backend endpoints, no new schemas, no Alembic migrations. - No widening of agent_require_approval (save_scenario already listed; HITL step consumes it). - No CRLF/LF line-ending normalisation bundled in. Contract probe report: PRPs/ai_docs/prp-41-contract-probe-report.md
…lish feat(api,ui): showcase pipeline agent ops final polish (PRP-41)
…tifact-uri fix(api): repair showcase safer promote cascade
![sourcery-ai[bot]](https://avatars.githubusercontent.com/in/48477?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Sorry @w7-mgfcode, your pull request is larger than the review limit of 150000 diff characters
|
Important Review skippedToo many files! This PR contains 152 files, which is 2 over the limit of 150. To get a review, narrow the scope: ⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: ⛔ Files ignored due to path filters (1)
📒 Files selected for processing (152)
You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Promotes the #324
showcase_richsafer-promote → scenario-replay cascade fix and the showcase manual demo guide fromdevtomainfor the next release cut (v0.2.20).devis 41 commits ahead ofmain; this PR cuts the accumulated work, with the #324 release-blocker as the headline.Key landed work
ctx.winning_run_idinstead of the swappeddemo-productionalias.safer_promote_flowwrites real-shape, parseableartifact_urivalues (no more unparseable placeholder)._restore_demo_alias_after_failure) added on pipeline failure sodemo-productionis never left on the worse-WAPE run.scenario_simulate_and_saveE2E tolerance removed — the step must now pass.docs/user-guide/showcase-manual-demo-guide.md).Validation
ruff check— passedruff format --check— passedpytest -m "not integration") — passedtest_run_demo_showcase_rich_full_epic— passed (live docker Postgres)mypy app/— only pre-existing xgboost stub baseline errors (none in changed files)Issue
Closes #324 after merge to
main/ the release flow (GitHub auto-closes on merge to the default branch).Merge this PR via the GitHub web UI — NOT
gh pr merge --merge— to avoid the release-please merge-commit-subject trap (a conventional non-bumping subject makes release-please skip the bump). Seedocs/_base/RUNBOOKS.md§ "release-please skipped the bump after a dev → main merge".