Model/docs/HANDOVER_MODELLING.md

# HANDOVER — Modelling stage rebuild

**Branch:** `feature/bill-derivation` (worktree `/workspaces/home/hestia-worktrees/model-assemble-new-backend`). **HEAD:** `6f0dcc04`.
**PRD:** GitHub `Hestia-Homes/Model#1152`, sliced into #1153–#1161. **All slices #1153–#1161 closed.**

## Issue status

| Issue | What | State |
|---|---|---|
| #1153 | Overlay Applicator + `EpcSimulation` | ✅ closed |
| #1154 | Package Scorer | ✅ closed — Elmhurst cascade pin (`4c0a907a`) |
| #1155 | wall Recommendation Generator | ✅ closed; cascade-pinned |
| #1156 | score Options + attribution | ✅ closed |
| #1157 | persist a Plan via `ModellingOrchestrator` | ✅ **closed this session** (`772cdd4f`→`c7e2aa37`) |
| #1158 | roof (loft) generator | ✅ closed — 300 mm + cascade pin |
| #1159 | floor generator | ✅ closed — overlay insulation-type field + pins |
| #1160 | Optimiser (knapsack + greedy repair) | ✅ **closed this session** (`77983cae`→`34d4748a`) |
| #1161 | Measure Dependency (ventilation) | ✅ **closed this session** (`7c59e919`→`0fec0699`) |

## What this session did

1. **Cascade pins for #1154/#1158/#1159** — `tests/domain/modelling/test_elmhurst_cascade_pins.py`. Parse Elmhurst before/after recommendation Summaries via the extractor chain (NOT `parse_site_notes_pdf`), apply the generator's overlay, score, assert delta 0 vs the after-cert. Found+fixed: loft 270→**300** mm; suspended floor needs the overlay to also set `floor_insulation_type_str='Retro-fitted'`.
2. **`ProductJsonRepository`** (`cc0bb8f9`) — file-backed catalogue behind the `ProductRepository` port.
3. **#1157 — persist a Plan.** Design review (`/grill-with-docs`) + 5 TDD slices. See "Design decisions" below.
4. **#1160 — the Optimiser.** 4 TDD slices. See "Design decisions".

## Design decisions locked this session (READ THESE)

- **Multi-phase is DEFERRED** (speculative prospective-client ask). **ADR-0005 rewritten to "Deferred".** No `plan_phase` table, no `phase` column. `CONTEXT.md` no longer has Scenario Phase / Plan Phase / Rolled-over Options. Everything is **single-phase**. Future: a migration adds `plan_phase` + back-fills live plans as 1-phase.
- **Plan Measure** is the new term (in `CONTEXT.md`): the persisted selected Option + its role-3 attributed impact + cost. **Recommendation** stays the *candidate* (never persisted; no stored impact).
- **Reuse the LIVE tables** (`plan`, `recommendation`) — they exist in the live product (`backend/app/db/models/recommendations.py`, SQLAlchemy `Base`) and the FE reads them. The rebuild writes the **same physical tables via SQLModel mirrors** (`infrastructure/postgres/plan_table.py`) — the established pattern (`task_table.py`→`tasks`, `product_table.py`→`material`). **ADR-0017** records this.
- Added **`recommendation.plan_id`** (FK→plan, ON DELETE CASCADE); retire the `plan_recommendations` m2m for new writes. FE-owned Drizzle migration: `docs/migrations/recommendation-plan-id.md`.
- Tracer persists **SAP + CO₂ (tonnes = calc kg ÷ 1000) + cost + derived `post_epc_rating`**. Energy/bill columns deferred. Idempotent replace per (property_id, scenario_id).
- **Optimiser = exact pure-Python multiple-choice knapsack**, NOT `mip`. Recycles `GainOptimiser`/`CostOptimiser`'s *formulation* (≤1/group, maximise gain s.t. budget) but not the dependency — **`mip`'s CBC backend does not load on this aarch64 container** (`NameError: cbclib`), so the legacy solver can't run/be tested here. ADR-0016's MILP is only a warm-start signal, so exact small-scale enumeration is ample. Re-score + greedy-repair toward the goal's SAP target gives the truth.

## `domain/modelling/` layout (grouped `84ec6da0`)

Behaviour lives in subpackages; shared value-object vocabulary stays flat at the top (imported everywhere): `recommendation.py` (Recommendation / MeasureOption / Cost), `plan.py`, `scenario.py`, `product.py`, `contingencies.py`, `simulation.py` (EpcSimulation overlay).
- `generators/` — `wall_recommendation` / `roof_recommendation` / `floor_recommendation`.
- `scoring/` — `overlay_applicator` (apply_simulations), `package_scorer` (role 2), `scoring` (role-1 `independent_option_impacts` + role-3 `marginal_impacts`). Note the path is `domain.modelling.scoring.scoring` for the role-1/3 module.
- `optimisation/` — `optimiser`, `measure_dependency`.

## What's built (all in `domain/modelling/`, `infrastructure/postgres/`, `repositories/`, `orchestration/`)

- Generators (`generators/`): `recommend_cavity_wall` / `recommend_loft_insulation` (300 mm) / `recommend_floor_insulation` (sets `floor_insulation_type_str`).
- `simulation.py` overlay + `scoring/overlay_applicator.apply_simulations` (generic field-fold) + `scoring/package_scorer.PackageScorer.score` (role 2) + `scoring/scoring.py` (`marginal_impacts` role 3, `independent_option_impacts` role 1).
- `scenario.py` `Scenario(id, goal, goal_value, budget, is_default)`; `plan.py` `Plan` + `PlanMeasure` (derives cost_of_works/contingency_cost/co2_savings/post_epc_rating).
- `optimisation/optimiser.py` — `optimise(groups, budget)` (exact knapsack) + `optimise_package(...)` (re-score + greedy repair, `Scorer` Protocol, `OptimisedPackage`).
- `infrastructure/postgres/`: `scenario_table.ScenarioRow`, `plan_table.{PlanRow,RecommendationRow}` (mirrors of live tables; `from_domain`).
- `repositories/`: `scenario/`, `plan/`, `product/` (Postgres + Json) — all on the `UnitOfWork` (`uow.scenario`/`uow.product`/`uow.plan`).
- `ModellingOrchestrator.run(property_ids, scenario_ids, portfolio_id)` — one UoW, commit once; generate (wall/roof/floor) → role-1 score → `optimise_package` → role-3 attribute → persist. Wired into `AraFirstRunPipeline` + `handler.py`.
- `datatypes/epc/domain/epc.py::Epc.sap_lower_bound()` (band → min SAP, target for INCREASING_EPC).

## Gotchas (will bite a fresh agent)

- **`mip` / CBC is broken on aarch64** here — never build runnable code on `mip`. The legacy `recommendations/optimiser/` tests only "pass" because they avoid constructing a `mip.Model`.
- **`moto` is not installed** — `tests/orchestration/test_postcode_splitter_orchestrator.py` and `tests/repositories/unstandardised_address/` fail at *collection*. Pre-existing, unrelated; `--ignore` them when sweeping.
- **Run tests:** `python -m pytest <path> -q` (do NOT pass `-p no:cov`). Ephemeral Postgres via the `db_engine` fixture builds **only `SQLModel.metadata`** — legacy `Base` tables are absent in tests, which is why mirrors work.
- **Worktree import trap:** `python /tmp/foo.py` imports `/workspaces/model`, not this worktree. Use `pytest` (rootdir handles it) or a `python -c` from the worktree root.
- **Driving Modelling in an integration test:** the calculator fixtures (`_elmhurst_worksheet_000490.build_epc()`) lack lodged recorded-performance fields, so the **Baseline stage can't run on them**. Drive `ModellingOrchestrator` directly off a repo-seeded EPC (`EpcPostgresRepository(session).save(epc, property_id, portfolio_id)`) — see `test_modelling_optimises_and_persists_a_multi_measure_plan`. The sample API EPC (`_lodged_epc()`) does go through the full pipeline.
- **`PortfolioGoal.INCREASING_EPC` value is `"Increasing EPC"`** (with a space) — the orchestrator compares `scenario.goal == "Increasing EPC"`.
- A generator calls `products.get(...)` during candidate generation, so the integration test must **seed a `material` row for every measure type that fires** (e.g. the sample EPC's uninsulated solid floor needs `solid_floor_insulation`).
- **Don't edit the SAP calculator's `heat_transmission.py`** (another agent owns it).

## Conventions

Commit per TDD slice; conventional-commit message ending `Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>`; stay on `feature/bill-derivation`. Tests use literal `# Arrange / # Act / # Assert`; assert with `abs(x - y) <= tol` (not `pytest.approx`); pyright strict, zero errors; annotate call-return locals. Cascade pins target the worksheet at delta 0.

## #1161 — Measure Dependency (ventilation), as built (4 TDD slices, all green)

Forks resolved with the user (AskUserQuestion): **guard now** (skip when already MEV/MVHR), **persist as a Plan Measure** (cost + real negative marginal), **forced but its cost counts toward spend** (mandatory-when-triggered, never budget-gated; repair sees less headroom).

1. **`7c59e919`** — Simulation Overlay grows a dwelling-level segment: `VentilationOverlay` (all-optional partial of `SapVentilation`, field `mechanical_ventilation_kind`) + `EpcSimulation.ventilation`; `apply_simulations` folds it onto `sap_ventilation` (creating one if the baseline lodged none). Until now the overlay was building-part only — ventilation is whole-dwelling.
2. **`6b11c902`** — generic injection in the optimiser: `MeasureDependency(triggers: frozenset[str], required: ScoredOption)` lives in `optimisation/optimiser.py` (its input contract). `optimise_package(..., dependencies=())` injects any dependency whose triggers ∩ selected-measure-types, before every re-score (initial **and** each repair). `_inject` dedups by required measure-type. Forced (injected even over budget) but its cost is in `_package_cost`, so repair headroom shrinks. `_best_repair_candidate` folds in any dependency a candidate newly triggers, so its marginal SAP and incremental cost are truthful; affordability gates on whole-package cost vs budget. Returned `selected` includes the injected deps. Optimiser stays domain-agnostic — no ventilation import.
3. **`1bf5b410`** — `domain/modelling/optimisation/measure_dependency.py`: `MEASURES_NEEDING_VENTILATION` (cavity/internal/external wall, cf. legacy `assumptions.measures_needing_ventilation`) + `ventilation_dependency(epc, products)` → MEV Option (`mechanical_ventilation_kind="EXTRACT_OR_PIV_OUTSIDE"`, decentralised MEV = legacy "mechanical, extract only"), priced at 2 fully-loaded units. Returns **None** when `sap_ventilation.mechanical_ventilation_kind` is already set (= legacy `has_ventilation` — confirmed against `backend/Property.py:1236`). Note: builder fetches the Product up-front, so the catalogue needs a `mechanical_ventilation` row for **every** not-yet-ventilated dwelling, even if no wall is ultimately selected.
4. **`0fec0699`** — orchestrator wiring: `_measure_dependencies` builds the (≤1) dependency; `_BEST_PRACTICE_ORDER` gains `"mechanical_ventilation"` between loft and floors (role-3 cascade walls→roof→**vent**→floor); ventilation persists as a Plan Measure with its real negative marginal + cost. Added `mechanical_ventilation: 0.26` contingency (legacy `Costs.CONTINGENCIES`). On 000490 the real calculator scores MEV at **−1.275 SAP**.

**Post-#1161 refactor (`631df921`→`02afc04c`):** production split from selection-semantics. Detection + pricing moved into a proper generator `generators/ventilation_recommendation.py::recommend_ventilation(epc, products) -> Optional[Recommendation]` (same shape as wall/roof/floor; guard returns None when already mechanically ventilated). `optimisation/measure_dependency.py` now owns only the trigger set + the forced-edge wrapping: `ventilation_dependency` delegates to the generator and wraps the Recommendation (cheapest Option) into the `MeasureDependency`. The orchestrator's `_measure_dependencies` call is unchanged. **Key asymmetry:** `recommend_ventilation` lives in `generators/` but is **not** in `_candidate_recommendations`' generator tuple — it's consumed only by the dependency path, never the free pool. This is the natural home for the multi-option future (MEV-c / MVHR) and the FE swap-in front.

Gotchas for the next agent: the ventilation Product/contingency must exist for any not-yet-ventilated dwelling (the generator fetches the Product at build time, not inject-time); the stub scorer in `test_optimiser.py` indexes `building_parts[MAIN]`, so vent-only overlays need the separate `_VentStubScorer`.

## Optimiser objective realigned to least-cost-to-target (`5620f49f`→`641c1bd7`)

A `/grill-with-docs` pass found the rebuild had the **wrong optimiser objective**: it maximised SAP gain within budget (target as a repair floor), whereas the legacy `StrategicOptimiser.solve()` Case 1 (the intended behaviour) is **min-cost subject to gain ≥ target and cost ≤ budget, fall back to max-gain only if the target is unreachable**. ADR-0016 was amended (it had specified the wrong objective). 4 slices, all green:

- **`05a4f5f8`** — `optimise_min_cost(groups, budget, target_gain, dependencies=())`: exact-enumeration sibling to `optimise`; cheapest package reaching `target_gain` within budget (ties → higher gain), `None` if unreachable.
- **`2bf42d04`** — `optimise_package` rewired: target present → min-cost warm-start → inject → re-score → repair toward target; if warm-start infeasible or repaired package still short on the true score → `_max_gain_package` fallback. No target → max-gain (unchanged). Stops at the target, no overshoot into a higher band, surplus budget unspent.
- **`af501fce`** — **ventilation-aware selection**: `_with_role1_signals` scores each dependency's true (negative) role-1 impact (was a `0.0` placeholder); `_augmented_cost_gain` folds the triggered dependency into every candidate's cost+gain in both selectors. Stops min-cost picking a wall whose mandatory ventilation (−1 to −5 SAP) it can't justify, or whose £900 a wall-free package would avoid.
- **`641c1bd7`** — orchestrator needed **no change** (already threads budget/target/deps); added an end-to-end pin (band-D property + goal D = already met → Plan with no measures).

Decisions locked (in the ADR amendment): target predicate `sap_continuous ≥ band_floor` (e.g. ≥ 69 for C — conservative, no legacy `allow_slack`); **budget is a hard envelope** — a wall whose ventilation would bust the budget is **dropped, not forced over** (reverses the earlier "forced regardless of budget" call; presence still guaranteed for any *selected* wall); warm-start-on-signal + re-score + repair kept (not exhaustive re-score) for scalability; "recommend slightly more than land short" is satisfied by the conservative floor + repair, not by spending budget for headroom.

## Bill-Derivation: plan-level post-retrofit bills (`75ba5dd7`→`198122d1`)

A `/grill-with-docs` pass designed the Modelling Bill-Derivation slice (ADR-0014 amended). Plan-level columns done across 4 slices; per-measure is the next slice.

- **`ced6287b`** — relocated `Bill` / `EnergyBreakdown` / `BillDerivation` / `sap_fuel` (+ tests) from `domain/property_baseline/` to a neutral **`domain/billing/`** (cross-stage concern; both Baseline and Modelling consume it). Pure move, ~10 files.
- **`2bbc401f`** — `Score` gains `sap_result: Optional[SapResult]`, populated by `PackageScorer`. Lets Modelling bill the scored end-state reusing a `SapResult` the optimiser/orchestrator already computed — **no second `calculate`**. Optimiser ignores it (stays `Score`-only; stubs unaffected).
- **`26de28aa`** — `Plan` carries optional `baseline_bill` / `post_bill` and derives `post_energy_bill` / `energy_bill_savings` / `post_energy_consumption` / `energy_consumption_savings` (None until billed → NULL).
- **`198122d1`** — `ModellingOrchestrator` gains a constructor-injected `FuelRatesRepository` (mirrors Baseline — `get_current()` once, one `BillDerivation` per batch); `_plan_for` bills the baseline (`scorer.score(epc, [])`) and post-package (`package.score`) `SapResult`s at the same snapshot, savings = baseline − post. `PlanRow` mirror + `from_domain` persist the four columns (they already exist on the live `plan` table — no FE migration). Pipeline/handler wired.

Key properties: **fuel-switch is handled for free** — we bill the fully-overlaid post-package `SapResult`, so a future oil→ASHP measure prices at the new fuel via `sap_code_to_fuel` (no per-measure fuel bookkeeping). Baseline and post are priced at one `FuelRates` snapshot, so the delta is rate-consistent. Carries ADR-0014's **appliances+cooking-stubbed-at-0** limitation (shared with Baseline, so savings stay consistent).

## Bill-Derivation: per-measure bill savings (`e79ffabf`→`b976c3ab`) — DONE

Filled `recommendation.kwh_savings` + `energy_cost_savings` via the **telescoping bill cascade** over the role-3 best-practice order. 3 slices, all green + pyright-strict-clean:

- **`e79ffabf`** — enabling refactor: pulled the cumulative-prefix scoring out of `marginal_impacts` into a reusable `scoring.cascade_scores(scorer, baseline, overlays) -> list[Score]` (index 0 = baseline, one `calculate` per prefix) + a pure `marginals_from_scores`. Each Score carries its `SapResult`, so the bill cascade re-bills the same prefixes the role-3 attribution scores — **no extra `calculate`**. `marginal_impacts` now delegates (behaviour unchanged).
- **`7e79c30a`** — `PlanMeasure` grows optional `kwh_savings` (delivered energy) + `energy_cost_savings` (£), signed so positive = saving, `None` until billed. `RecommendationRow` declares the live `recommendation.kwh_savings`/`energy_cost_savings` columns + maps them (None→NULL). Vestigial `recommendation.energy_savings` stays **undeclared** (legacy = 0). No FE migration (columns already live).
- **`b976c3ab`** — `_plan_for` scores baseline + every prefix once via `cascade_scores`, bills each at one Fuel Rates snapshot, and takes **consecutive Bill deltas** as each measure's marginal delivered-kWh + £ saving. The Plan's `baseline_bill`/`post_bill` are now the **same cascade endpoints** (`bills[0]`/`bills[-1]`), so per-measure savings telescope **exactly** to the headline savings — pinned on the real calculator (Σ per-measure == plan totals, abs ≤ 1e-6). Ventilation's saving is **negative** and still telescopes. Added `Bill.total_consumption_kwh` (shared by Plan + orchestrator); dropped the redundant standalone baseline `calculate`.

Key property: `MeasureImpact.energy_savings_kwh_per_yr` is *primary* energy and does **not** feed `kwh_savings` — `kwh_savings` is **delivered** energy from the Bill section kWh. Carries ADR-0014's appliances+cooking-stubbed-at-0 limitation.

## Retire `plan_recommendations` + consolidate models (`b76d0f81`→`6f0dcc04`) — DONE

Designed in `/grill-with-docs` + `/grill-me`. The live `plan`/`recommendation` tables are read **directly by the Drizzle FE**, so this was a two-repo expand/contract. **FE-visibility goal met:** Plans and their measures now link solely by `recommendation.plan_id`; the m2m is gone. 9 slices, all green + pyright-strict-clean, and the rebuild + legacy suites are now **co-runnable** (the consolidation fixed a pre-existing dual-definition collision).

- **`b76d0f81`** — migration spec ([docs/migrations/recommendation-plan-id.md](migrations/recommendation-plan-id.md): add `plan_id` → backfill → dual-write → cut reads → drop; backfill-before-reads + dual-write are the load-bearing rules since the FE can't deploy atomically) + ADR-0017 amendment.
- **`c1c7b06f`** — consolidate `plan`/`recommendation`/`recommendation_materials` into **`infrastructure/postgres/modelling/`** as single SQLModel defs (absorbing the partial `PlanRow`/`RecommendationRow` mirrors, full column parity + `plan_id`). `backend/app/db/models/recommendations.py` → re-export shim. Export conftest: create SQLModel-first / skip the redundant `drop_all` (the `epc` enum type is now shared across both metadatas).
- **`27fcc5b1`** — legacy writers set `recommendation.plan_id` (dual-write).
- **`af5dbe32`** — cut all three readers (`portfolio_functions`, `Outputs`, `export/property_scenarios`) onto `plan_id`.
- **`b97d0688`** — drop the m2m: writes, `delete_property_batch` cleanup, the `PlanRecommendationRow` model, the `test_export` fixtures.
- **`01c2c391`** — rename the cluster `…Row` → **`…Model`** (matches the `epc_property` precedent + the legacy names `backend/` already imports, so the shim's plan re-export is literal). The non-cluster `…Row` tables stay until their live legacy `…Model` counterparts retire (renaming now would re-create dual-definition collisions).
- **`2fbd7147`** — move `PortfolioGoal` to **`domain/modelling/portfolio_goal.py`** (domain vocab; infra→domain is the normal direction); `portfolio.py` keeps a re-export.
- **`c18968ba`** — consolidate `scenario` + `installed_measure` (full-parity `ScenarioModel`/`InstalledMeasureModel` + `MeasureType`). **`ScenarioModel.goal` is the `PortfolioGoal` enum** (legacy planning branches on it); the repo's `to_domain` maps it to its value, so `Scenario.goal` is now the value `"Increasing EPC"` consistent with the orchestrator — fixing the latent name-vs-value bug the old `str` column masked.
- **`6f0dcc04`** — characterization test for the FE aggregation `aggregate_portfolio_recommendations` (was untested), pinning the `plan_id` join.

**Gotchas for the next agent:** the modelling SQLModel classes are `…Model` and live in `infrastructure/postgres/modelling/` (NOT the old flat `plan_table.py`/`scenario_table.py` — deleted); `backend/app/db/models/recommendations.py` is now a pure shim. Out-of-cluster columns are plain ints (no FK) per the mirror convention. **`PortfolioGoal` lives in `domain/modelling/`** now. The `etl/`+`sfr/` reporting scripts still reference the m2m and are **deferred** (out of scope). The live DB changes (add `plan_id`, backfill, drop `plan_recommendations`) are the **FE-owned Drizzle** migrations in the migration doc — this branch is the backend end-state.

## NEXT PHASE — depth + scale e2e (handover for a grilling session)

The owner's goal: run a large dump of **SAP 10.2 EPCs (1,000–10,000)** through Modelling and inspect the recommendations — a large-scale integration test — plus **manual testing via a Python console**. Measure *coverage* (heating/solar/glazing/…) is explicitly **deferred** ("we'll flesh this out"). This phase is **depth + scale on the existing 4 fabric measures** (cavity wall / loft / floor / ventilation):

1. **Close the persisted-field gaps** so a persisted Plan matches the engine's richness for the measures we *do* model: `recommendation_materials` (BOM — depth/quantity/unit/cost; rebuild `Cost` is a single total today, no per-material breakdown), per-measure U-values (`starting_u_value`/`new_u_value`), `total_work_hours`/`labour_days`. Source of truth: the rebuild `ProductRepository` (`repositories/product/`) + legacy `materials_functions.py` / `recommendations_functions.upload_recommendations` (writes `rec["parts"]`).
2. **Financial uplift modelling** — valuation columns (`plan.valuation_*`, `recommendation.property_valuation_increase`/`rental_yield_increase`) are **greenfield in the rebuild** (no domain concept yet). Legacy logic: `backend/Property.py`, `backend/Funding.py`, `backend/app/db/functions/funding_functions.py`, `portfolio_functions.py`. Needs a domain design (likely a `/grill-with-docs` pass).
3. **Large-scale e2e harness** — template is `tests/orchestration/test_ara_first_run_pipeline_integration.py::test_modelling_optimises_and_persists_a_multi_measure_plan` (seeds an EPC via `EpcPostgresRepository` + `MaterialRow`s + a `ScenarioModel`, runs `ModellingOrchestrator` directly — the Baseline stage can't run on calculator fixtures). For the dump: parse each EPC via `EpcPropertyDataMapper.from_api_response` / `from_rdsap_schema_21_0_x` (see `datatypes/epc/domain/mapper.py`), seed, run, inspect. EPC samples live under `backend/epc_api/json_samples/`.
4. **Python-console manual run** — instantiate `ModellingOrchestrator` against a real DB and inspect Plans/Recommendations. Mind the **worktree import trap** (run from the worktree root, not `/tmp`).

A self-contained handover prompt for the next agent is in **`docs/HANDOVER_NEXT_PHASE_PROMPT.md`**.

## What's left

**Deferred fronts** (open, post-#1161): exclusion-filtering of the candidate pool (deferred from #1160); persist **unselected alternatives** (`default=False` rows linked via `plan_id`) for the swap-in UX — open ADR-0016 question: what impact figure they carry; promote `ProductRepository` to the DB+file composite; non-EPC goal objectives (Energy Savings, Reducing CO2) in the optimiser. Possible extension of the ventilation trigger set to roof insulation (now a one-line data edit in `MEASURES_NEEDING_VENTILATION`); and making the dependency builder lazy (thunk) so the Product is only fetched when a trigger is actually selected.

## Key references

- ADRs: **0005** (multi-phase deferred), **0011/0012** (orchestrators + UoW), **0016** (three scoring roles + warm-start/re-score/repair), **0017** (Plan persistence — evolve live tables).
- `CONTEXT.md`: Plan, Plan Measure, Recommendation, Measure Option, Optimised Package, Scenario, Measure Dependency.
- Auto-memory `project_modelling_stage_state` has the running state.