VSS: Value of the Stochastic Solution report (--vss, two-stage)

[DLWoodruff]

What

Adds a Value of the Stochastic Solution (VSS) report to
generic_cylinders. Passing --vss prints RP / EV / EEV / VSS at the
end of an --EF or a decomposition run. Two-stage, opt-in, report-only
(it never changes the solution). Closes the two-stage part of the
pysp_but_not_mpisppy.md §A7 gap (PySP's ef_vss.py).

Example (farmer, 3 scenarios):

================= VSS report =================
  RP  (stochastic solution, here-and-now):        -108390   [EF, exact]
  EV  (mean-value problem objective)     :        -118600
  EEV (EV first stage over scenarios)    :        -107240
  VSS = EEV - RP (sense=min)             :           1150   (1.06% of |RP|)
=============================================

How

Mirrors the do_mmw post-run report and reuses shipped machinery — no
new spoke/hub code:

• average_scenario_creator (the Jensen's-bound contract) → the mean-value
  solve, giving EV and the first-stage x_bar;
• Xhat_Eval.evaluate → the honest cross-scenario EEV (fix x_bar,
  solve every scenario);
• RP from ef.get_objective_value() — exact when the EF is solved to a
  zero gap (always for an LP), otherwise the incumbent bracketed by the
  solver's dual bound; or, from a decomposition run, the incumbent
  (BestInnerBound) with a BestOuterBound bracket from the wheel.

New mpisppy/generic/vss.py (vss_prep + do_vss), one config flag, two
call sites in generic_cylinders.py.

Behavior details

• Sign-aware: VSS = EEV − RP (min) or RP − EEV (max), always
  reported ≥ 0.
• EF RP is gap-aware: an --EF run reports RP as [EF, exact] only
  when the solver reaches a zero gap (an LP, or a MIP solved tight). A MIP
  left with a nonzero mipgap returns only an incumbent, so RP is relabelled
  [EF incumbent, gap=…] and bracketed by the solver's dual bound — the same
  point-plus-bracket treatment as the decomposition case below. (do_EF
  stashes the EF objective bound for this; a solver that reports no bound
  falls back to [EF, exact].)
• Decomposition: conservative point value (against the incumbent) plus
  the VSS bracket implied by any unclosed optimality gap.
• Solver options / mipgap: the EV and EEV solves reuse the run's solver
  and solver_options via solver_specification — EF_solver_options
  after --EF, else solver_options — so all three numbers are solved
  consistently and any mipgap in that string applies. Since VSS = EEV − RP
  is a difference of two optimized values, a loose gap makes a small VSS
  unreliable (documented). (Also lets --vss work on an --EF run that
  sets only --EF-solver-name.)
• Infeasible mean-value first stage (no relatively complete recourse) ⇒
  EEV = VSS = +inf, naming the offending scenarios. MPI-safe: solves with
  need_solution=False so a clean per-scenario infeasibility does not
  deadlock the collectives.
• Fail-fast at setup (before the solve) on multistage, proper bundles,
  ADMM, --cvar, or a missing average_scenario_creator.

Cost

EEV re-solves every scenario once with the first stage fixed. Fixing the
first stage decouples the scenarios and makes each solve easier than the
original, so the added wall-clock is model-dependent — usually a fraction
of the run (for a decomposition run, ~one iteration's worth), occasionally
significant with very many scenarios or expensive recourse. Off by default.

Tests

mpisppy/tests/test_vss.py: setup guards, sign/bracket helpers, EF
minimize and maximize end-to-end (asserting the actual RP/EV/EEV/VSS
numbers), solver-options threading, the EF RP labelling (exact / nonzero-gap
bracket / zero-gap) and its dual-bound helpers, the decomposition bracket,
and the infeasible-EEV path. Wired into run_coverage.bash and a dedicated step in
the solver-equipped regression CI job so the numerical assertions run
under cplex/xpress (not just the no-solver unit job).

Drive-by: test_jensens.py had the same wiring gap (only in the
no-solver job); added a regression-job step so its solver-gated e2e cases
run in CI too.

Docs

doc/src/vss.rst (in the toctree; explains VSS, the honest cost note, the
solver-options/mipgap reuse, the EF-exact vs bracketed-RP distinction, and
infeasibility); doc/designs/vss_design.md (design);
pysp_but_not_mpisppy.md §A7 marked addressed for two-stage.

Scope / follow-ups

Two-stage only. Multistage VSS (the expected-value tree via
ef_xhat_nonants) and EVPI are noted as future work in the design doc. V1
threads the solver-options string (not the separate EF_mipgap /
*_iter*_mipgap knobs or a *_solver_options_file).

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 84.97110% with 26 lines in your changes missing coverage. Please review.
✅ Project coverage is 76.24%. Comparing base (b2474b4) to head (e3664d0).
⚠️ Report is 6 commits behind head on main.

Files with missing lines	Patch %	Lines
mpisppy/generic/vss.py	88.46%	18 Missing ⚠️
mpisppy/generic_cylinders.py	40.00%	6 Missing ⚠️
mpisppy/generic/ef.py	50.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #791      +/-   ##
==========================================
+ Coverage   76.16%   76.24%   +0.07%     
==========================================
  Files         169      170       +1     
  Lines       22224    22396     +172     
==========================================
+ Hits        16928    17076     +148     
- Misses       5296     5320      +24     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.