0.5.0 - 2026-04-25

Release Notes

Major refactor. Consumers must update config.json, any code calling
solver traits, and any tooling reading solver/iterations.parquet
before upgrading.

Breaking Changes

Public Rust API:

• TrainingResult is #[non_exhaustive]; use TrainingResult::new(...).
• SolverInterface::record_padding_stats → record_reconstruction_stats(&mut self).
• SolverInterface::clear_solver_state removed.
• WarmStartBasisMode and CanonicalStateStrategy enums removed along
  with the canonical_state_strategy config fields.
• Four classification counters (basis_preserved, basis_new_tight,
  basis_new_slack, basis_demotions) collapsed into basis_reconstructions.
• basis_rejections + basis_non_alien_rejections → basis_consistency_failures.
• add_rows_count, add_rows_time_ms, clear_solver_count, and
  clear_solver_failures counters removed.
• StoppingRuleResult.rule_name is now &'static str;
  StoppingRuleResult.detail is now Cow<'static, str>. Call
  .to_string() on rule_name or use detail.as_ref() at consumers
  that previously expected String.
• TrainingEvent::WorkerTiming.timings field type changed from
  [f64; 16] (where 12 of 16 slots were always zero on per-worker
  events) to a new WorkerPhaseTimings struct with four named fields:
  forward_wall_ms, backward_wall_ms, fwd_setup_ms, bwd_setup_ms.
  The output Parquet schema for training/timing/iterations.parquet
  is unchanged. Consumers that read the variant payload directly must
  access the named fields rather than slot indices. The rank-aggregated
  WorkerTimingRecord writer record retains its 16-column layout and
  the WORKER_TIMING_SLOT_* constants remain public as the bridge
  between named fields and writer slots.

Output schema — solver/iterations.parquet drops from 23 to 19
columns. Added: opening, rank, worker_id (nullable i32),
basis_reconstructions (u64). Removed: the eight counters listed
above. Backward rows gain an opening dimension; forward rows are now
one per (iteration, stage) rather than one per iteration.

cut_selection/iterations.parquet drops active_after_angular
(10 → 9 columns).

Policy FlatBuffer — CUT_FIELD_DOMINATION_COUNT slot removed. Old
policies deserialise via graceful-absence; "domination_count"
disappears from cobre.results.load_policy per-cut dicts.

Added

• Backward-pass basis cache — rank 0 captures a fresh basis per
  stage during the backward pass and broadcasts it end-of-iteration;
  next-iteration backward solves warm-start from the cache.
• Basis reconstruction — CapturedBasis wrapper with slot and
  state metadata; reconstruct_basis applies a stored basis across
  cut-set churn on forward, backward, and simulation paths. Controlled
  by training.cut_selection.basis_activity_window (1-31, default 5).
• Weekly+monthly studies — sub-monthly lag accumulation,
  recent_observations input for mid-season starts, terminal boundary
  cuts (policy.boundary.{path, source_stage}) for Cobre-to-Cobre FCF
  coupling, and non-uniform per-stage scenario counts.
• Multi-resolution studies — same-season noise group sharing,
  observation aggregation from finer to coarser resolution, and
  monthly→quarterly PAR transition.
• Per-opening solver statistics — backward rows carry rank and
  worker_id metadata for per-worker parity testing.
• TrainingEvent::SimulationStarted variant mirrors TrainingStarted
  for the simulation phase. Carries n_scenarios, n_stages, ranks,
  and threads_per_rank. Emitted once per rank before the parallel
  scenario loop so consumers can render a banner before any scenario
  completes.
• Non-TTY progress lines (pipes, mpirun aggregators, log files) now
  include an [elapsed HH:MM:SS < eta HH:MM:SS] trailing cell matching
  the interactive bar.
• Wire-format version bytes added to both the basis broadcast payload
  and cut records. The basis version field is at position 1 of the
  i32 metadata buffer; the cut version byte is at byte 0 of each
  serialized cut record. Mismatches between sender and receiver produce
  typed validation errors rather than silent data corruption or panics.
• Typed errors for MPI correctness conditions:
  • Non-uniform worker counts across ranks are detected during the
    backward-pass handshake and reported as a validation error naming
    both the minimum and maximum observed counts.
  • Basis broadcast i32 length overflow (more than i32::MAX
    elements) is detected before the MPI call and reported as
    CommError::InvalidBufferSize.
  • Wire-format version mismatches in the basis broadcast and in cut
    deserialization are reported as validation errors naming both
    expected and received versions.
  • Violated cut-count invariants in sync_cuts (negotiated count
    differs from actual serialized count) are reported as a validation
    error.
  • Stats counter values exceeding 2^53 (the f64 integer precision
    limit), which would corrupt MPI allreduce sums, are detected before
    the allreduce call and reported as an internal error.

Changed

• Lower-bound LP is strictly append-only. Cuts are never removed
  from the LB LP, guaranteeing monotonic non-decreasing lower bound.
  Cut selection and budget enforcement continue to deactivate cuts in
  the shared pool for the forward and backward passes.
• Cut management pipeline reduced to two stages: strategy-based
  selection followed by budget enforcement.
• Internal refactor of cobre-io semantic-validation: the
  6 319-line validation/semantic.rs file is now split into
  7 cohesive domain submodules (hydro, thermal, stages,
  scenarios, season, correlation, sobol) plus a
  placeholder shared module for future cross-domain helpers.
  The two public entry functions
  (validate_semantic_hydro_thermal,
  validate_semantic_stages_penalties_scenarios) keep their
  paths under cobre_io::validation::semantic::*. No semantic-
  validation rule was added, removed, or modified.
• TrainingEvent::SimulationProgress.scenarios_complete now carries a
  global estimate under multi-rank execution (local × ranks, clamped
  to scenarios_total), matching the scenarios_total field's global
  scope. Single-rank runs are unchanged. Previously, rank 0 reported
  its local count against the global total, producing a misleading
  50/100 final line on a 2-rank run.
• cobre-comm now requires ferrompi >= 0.4.0. The MPI bitwise-OR
  allreduce used to synchronize the active-window bitmap is dispatched
  directly to MPI_Allreduce with MPI_BOR instead of being emulated
  via an allgatherv + bytewise fold workaround.
• FPHA hyperplane file is now written to
  <output_dir>/hydro_models/fpha_hyperplanes.parquet in both the CLI
  (--output flag) and Python bindings (output_dir parameter).
  Previously the file was always written to
  <case_dir>/output/hydro_models/, ignoring the caller-specified
  output directory. Under multi-rank execution the file is written by
  rank 0 only; previously all ranks raced to write the same path.
• All sort comparators on f64 in the hot path now use f64::total_cmp
  for deterministic NaN handling. partial_cmp(...).unwrap_or(Equal)
  patterns on the production training and simulation paths have been
  replaced.
• Rayon global thread pool initialization failure now emits a structured
  warning (configured threads, actual threads, error reason) and
  continues using the already-initialized pool. Previously a silent
  fallback was applied without any diagnostic output.

Deprecated

• RowSelectionConfig::threshold is deprecated. The field remains
  parseable; supplying it now emits a WARN-level tracing event
  directing the user to memory_window (for "lml1") or
  domination_epsilon (for "domination").

Removed

• Angular diversity pruning — config section, module, output
  column, and training event variant.
• SparseCut module — handled by the indexer.
• CutMetadata::domination_count — unused.
• LB LP periodic rebuild and its CutRowMap helpers.
• FerrompiScratch and the FerrompiBackend::scratch interior-
  mutability field. The native bitwise-OR allreduce no longer needs
  per-call scratch for counts/displs vectors.
  unsafe impl Send + Sync for FerrompiBackend comments were updated
  accordingly.
• SimulationSummary::stage_stats: Option<Vec<StageSummaryStats>>
  field and the StageSummaryStats struct were removed; the field was
  always None in production output and had no consumers.
• ExportsConfig reduced to two flags: states and stochastic. The
  seven previously declared fields (training, cuts, vertices,
  simulation, forward_detail, backward_detail, compression)
  had no runtime consumers and have been removed from the public
  config surface. Existing config.json files that set these keys
  will continue to load — the keys are silently ignored.
• Cargo features tcp = [] and shm = [] removed from cobre-comm
  (no runtime backends were declared under either feature).
  highs = [] removed from cobre-solver; the HiGHS FFI is compiled
  unconditionally. Downstream Cargo.toml files that specified
  cobre-comm = { features = ["tcp"] } or
  cobre-solver = { features = ["highs"] } must drop those entries.
• cobre-sddp crate root re-export surface reduced from ~85 to ~50
  symbols. Implementation-detail types (BackwardOutcome,
  CutSyncBuffers, StageIndexer, FphaColumnLayout, EvapConfig,
  LbEvalScratch*, PatchBuffer, WorkspacePool, BasisStore*,
  CapturedBasis, RiskMeasure*, EvaporationModel*, FphaPlane,
  LinearizedEvaporation, MonitorState, TrainingOutcome,
  TrainingContext, StageContext, and similar) are no longer
  accessible at cobre_sddp::Type. They remain reachable via their
  full module path (e.g., cobre_sddp::cut::pool::CutPool,
  cobre_sddp::workspace::SolverWorkspace).

Verified

• D01-D30 + convertido SHA256 map matches v0.4.5 reference except for
  the documented schema renames. All 90 stable parquet entries are
  byte-identical versus the previous baseline.
• D01-D15 parity hash holds across all internal refactors landed in
  this release (cut coefficients, primal/dual trajectories,
  convergence trajectory bit-for-bit identical).

cobre-cli 0.5.0

Install cobre-cli 0.5.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/cobre-rs/cobre/releases/download/v0.5.0/cobre-cli-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/cobre-rs/cobre/releases/download/v0.5.0/cobre-cli-installer.ps1 | iex"

Download cobre-cli 0.5.0

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 0.5.0

Install cobre-mcp 0.5.0

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/cobre-rs/cobre/releases/download/v0.5.0/cobre-mcp-installer.sh | sh

Install prebuilt binaries via powershell script

powershell -ExecutionPolicy Bypass -c "irm https://github.com/cobre-rs/cobre/releases/download/v0.5.0/cobre-mcp-installer.ps1 | iex"

Download cobre-mcp 0.5.0

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