0.6.2 - 2026-05-20

Release Notes

Added

• New allow_curtailment: bool field on NonControllableSource (default
  true, preserving existing case behaviour). When set to false on an
  entity, the LP pins its generation column to the realized availability
  on every scenario: lower and upper column bounds both equal
  max_generation_mw * alpha * block_factor, where
  alpha = clamp(mean + std*eta, 0, 1) is the per-(stage, scenario)
  availability ratio drawn from non_controllable_stats.parquet and
  block_factor is the per-(stage, block) shape factor from
  non_controllable_factors.json. Use this on NEWAVE-derived
  geracao_usinas_nao_simuladas aggregates (PCH, PCT, EOL, UFV, MMGD)
  that the source model pre-nets from MERC; with the default Cobre
  schema the LP was free to curtail these because curtailment is one of
  the cheapest LP slacks, leading on the bundled deterministic 1983 case
  to ≈ 18 % of total NCS supply being curtailed, a ≈ +15 % hydro-dispatch
  swing, and ≈ −23 % spillage versus NEWAVE. Setting
  allow_curtailment = false on the must-run aggregates restores
  dispatch parity with NEWAVE while preserving per-source observability
  in the simulation outputs. JSON schema accepts the field as optional;
  absent → true.
• ModelProvenanceReport.historical_library_past_inflows_digest: Option<u64>
  records a SipHash-1-3 fingerprint of initial_conditions.past_inflows
  when the historical scenario scheme is active. Downstream consumers can
  compare the stored digest against the current past_inflows to detect
  a stale historical library that needs re-standardisation. Serialised
  under #[serde(skip_serializing_if)] so non-historical runs keep their
  JSON provenance unchanged. The cobre-python production path populates
  the field from setup.scenario_libraries.training.historical.

Changed

• API (breaking, hydro penalty rename): HydroPenalties.fpha_turbined_cost
  is renamed to HydroPenalties.turbined_cost, cascading through
  RawHydroPenalties, the penalty_overrides.parquet column, the output
  schemas, and the simulation_writer record. The turbined-flow
  regularisation cost is now applied universally to every hydro's turbine
  column in the LP objective (previously gated behind is_fpha in
  fill_turbine_columns, so constant-productivity plants paid nothing and
  diverged from NEWAVE). The dead turbined_cost field on
  ResolvedProductionModel::Fpha is dropped (the cost now lives in the
  penalty cascade). A latent extraction bug is fixed in the process:
  SimulationCostResult.turbined_cost was summing primal*obj over
  indexer.generation (FPHA generation columns, objective = 0), so the
  field was silently zero in real runs; it now sums over indexer.turbine
  where the cost lives. JSON schemas regenerated via cobre schema export
  with no hand edits. Two LB regression pins shift by ≈ 0.10 % under the
  new universal cost (basis_reconstruct_churn::PINNED_FINAL_LB,
  deterministic::D03_EXPECTED_COST).
• API (breaking, PAR primitives): dropped the explicit order: usize
  parameter from evaluate_par, evaluate_par_inflow, solve_par_noise, and
  removed the equivalent par.order(h) inner-loop bound from
  evaluate_par_batch / solve_par_noise_batch / solve_par_noises /
  evaluate_par_inflows. All four primitives now iterate the full
  psi.iter() slice (via psi.len() in the batch variants), and require
  lags.len() >= psi.len(). psi.len() is the authoritative number of lag
  terms in the model — equal to the AR order for classical PAR(p) and to 12
  when PAR(p)-A annual is active (the materialised annual coefficient is
  spread across the extra positions in PrecomputedPar::psi_slice). Callers
  that previously truncated to AR order silently dropped the annual
  contribution; see the matching Fixed entry below.

Fixed

• PAR(p)-A annual coefficients were omitted from the cut sparse mask,
  producing over-estimating Benders cuts and LB > UB at convergence.
  StudySetup::new was passing par.order(h) (the classical AR order) as
  the per-hydro lag-state-slot count to StageIndexer::set_nonzero_mask.
  With PAR(p)-A active, PrecomputedPar::psi_slice widens to
  max_order = 12 and the standardised annual coefficient ψ̂/12 fills
  positions p..12. The LP loads all 12 psi entries into the AR-dynamics
  row, but the cut sparse mask omitted state coefficients on lag slots
  p..12 — so cut rows built via build_cut_row_batch_into were missing
  the trailing state dependencies. At the visited state this shifted the
  cut hyperplane above the true LP value, producing systematically
  over-estimating cuts. On the bundled NEWAVE 1983 case (Camargos
  AR(4)+A on every hydro), the gap converged to LB > UB by ≈ 7 % with 41
  of 50 iterations recording negative gaps; disabling annual via
  order_selection = pacf collapsed the gap to 0.05 %. After the fix the
  same case with pacf_annual converges to 0.003 % final gap with zero
  negative-gap iterations across all 50. PrecomputedPar now records a
  per-hydro has_annual: Box<[bool]> at build time and exposes
  effective_lag_count(h) (returns max_order when has_annual[h],
  else orders[h]); StageIndexer::set_nonzero_mask's ar_orders
  parameter is renamed to lag_counts to reflect its true meaning. New
  regression test nonzero_mask_par_a_includes_full_psi_stride pins
  the contract. This is the direct analog of the η-inversion bug fixed
  above — same order-vs-psi.len() confusion, cut-construction path
  instead of standardisation path.

• PAR(p)-A annual was silently dropped at standardisation time when the
  scheme was historical or external. standardize_historical_windows
  (and the analogous external path) built their per-hydro lag_buf only up
  to par.order(h) lag slots and passed order_h as the iteration bound to
  solve_par_noise. With PAR-A active, PrecomputedPar::psi_slice returns
  12 entries (AR coefficients in slots 0..order, plus the spread annual
  coefficient ψ̂/12 in slots order..12), all of which the LP loads via
  lp_builder/matrix.rs:834-841. The standardised η therefore omitted the
  annual contribution from lags order..12, while the LP applied it in
  full — so forward replays produced biased inflows even when the stage-0
  lag state matched the window's pre-study lags exactly. Worst-case observed
  on the bundled NEWAVE 1983-anchored case (Camargos, AR(4) +
  annual_coefficient ≈ −0.225): a ≈ 11 % shortfall vs the raw historical
  observation. After the fix, the same case reconstructs every hydro at
  stage 0 to within 10⁻¹² of the historical observation. The fix combines
  the API change above with two callsite updates that now fill the full
  psi.len() lag slots in standardize_historical_windows and
  standardize_external_inflow. New regression test
  crates/cobre-stochastic/tests/par_a_historical_replay.rs pins the
  invariant directly.

• ClassSampler::Historical::apply_initial_state no longer overwrites the
  inflow-lag portion of the stage-0 state vector with the window-preceding
  raw historical inflows. Previously, the forward pass replayed the lag
  state of the historical year being sampled (e.g. when scenario m
  replayed 1983, the forward LP started from the 1982-Q4 lag values), while
  the lower-bound and backward-pass evaluators kept the user-supplied
  initial_conditions.past_inflows lags. The two paths therefore evaluated
  V₀ at different x_0 on every historical-replay case, producing a
  structural, typically negative SDDP gap (≈ −19 % on the bundled
  NEWAVE-derived 1983 deterministic case) that did not close with
  iteration count. The historical window now contributes only its
  standardized noise residuals via fill; the initial inflow lags come
  uniformly from initial_conditions.past_inflows for every scenario,
  matching NEWAVE's TENDENCIA HIDROLOGICA convention. Cases using
  InSample, OutOfSample, or External schemes are unaffected
  (bit-identical output — those variants were already no-ops). Cases using
  the Historical scheme will see different forward upper bounds and
  meaningful gap closure to cut tightness.

• Historical η inversion re-rooted at past_inflows (LB/UB x₀ sharing).
  standardize_historical_windows previously inverted the PAR noise residual
  η for each window stage using window_pre_study_lags as the lag state seed
  — i.e. the raw historical inflows from the year before each window year.
  After commit dc96030 (apply_initial_state no-op), the SDDP forward pass
  starts every scenario from initial_conditions.past_inflows (the
  user-supplied x₀), not from the window-preceding lags. The two paths
  therefore built their lag chains from different x₀, producing a
  systematic per-stage offset z_h = target + Σψ·(past_inflows − window_lag) that propagated through all stages and prevented exact
  historical replay even at stage 0 after d0e4a42. The fix re-roots the
  inversion on a rolling lag chain seeded from past_inflows, following the
  same accumulate/finalize pattern as standardize_external_inflow. The
  rolling chain is advanced each month via StageLagTransition; uniform-
  monthly transitions are used by default, with noop transitions (produced
  by an empty SeasonMap) falling back to uniform-monthly in the inner loop.
  A past_inflows_digest: u64 field (SipHash-1-3 of all past_inflows
  values) is stored on HistoricalScenarioLibrary to enable stale-library
  detection when past_inflows change between calls. A debug_assert!
  guard fires when stage_lag_transitions contain non-trivial (non-monthly)
  entries, marking TODO(historical-replay-non-monthly) for future work on
  sub-monthly and multi-monthly study configurations.
  build_historical_inflow_library and build_opening_tree_library both
  forward past_inflows and stage_lag_transitions to
  standardize_historical_windows. After the fix, every historical-scheme
  forward pass that starts from past_inflows reconstructs the raw
  historical observation to within floating-point precision at every stage,
  restoring LB/UB consistency at x₀. New tests T2–T6 in
  crates/cobre-stochastic/tests/par_a_historical_replay.rs pin the
  invariant for differing past_inflows, AR(0), truncated lag vectors,
  multi-window cases, and the non-trivial-transition guard.

cobre-cli 0.6.2

Install cobre-cli 0.6.2

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/cobre-rs/cobre/releases/download/v0.6.2/cobre-cli-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/cobre-rs/cobre/releases/download/v0.6.2/cobre-cli-installer.ps1 | iex"

Download cobre-cli 0.6.2

File	Platform	Checksum
cobre-cli-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
cobre-cli-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
cobre-cli-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
cobre-cli-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
cobre-cli-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

cobre-mcp 0.6.2

Install cobre-mcp 0.6.2

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/cobre-rs/cobre/releases/download/v0.6.2/cobre-mcp-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/cobre-rs/cobre/releases/download/v0.6.2/cobre-mcp-installer.ps1 | iex"

Download cobre-mcp 0.6.2

File	Platform	Checksum
cobre-mcp-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
cobre-mcp-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
cobre-mcp-x86_64-pc-windows-msvc.zip	x64 Windows	checksum
cobre-mcp-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
cobre-mcp-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum