Investment non-anticipativity window — the standard two-stage stochastic investment mode

[jkiviluo]

Stacked on the Slice G branch (top of the recourse stack). Delivers the standard two-stage stochastic capacity-expansion mode — the one that was previously not achievable: a single shared here-and-now investment over the first period(s), per-scenario recourse investment later, and operations free per scenario from timestep 0.

What it adds

New solve param non_anticipativity_invest_periods (Array of periods; schema v72). It ties v_invest/v_divest across stochastic branches over the chosen periods — a single shared decision where the branches share information.

• unset / [] (default) → no tie → byte-identical to today (recourse stays per-branch, none stays shared).
• [period1] under recourse → the main mode: investment shared in period 1 (the genuine here-and-now first stage), per-branch recourse investment afterward. Combine with Slice G non_anticipativity_periods = [] for operations free from t0.

This finally decouples the shared first stage from the operational tie: previously you could only get all-shared investment (none), all-per-branch (recourse), or the mid-horizon hedge (which also shares near-term operations). The investment-NA window consumes the pd_non_anticipativity frame that Slice A built and Slice E deferred, adding the non_anticipativity_invest/divest_p/n constraint families.

Correctness points (from the adversarial review)

• All four families are tied — invest and divest, process and node. Tying only v_invest would leak anticipativity through the net first-stage capacity (existing + invest − divest) for any reinvest/lifetime entity.
• The constraint's domain-guard join is cross-Enum (branch vocabulary vs period vocabulary); it uses the Utf8-cast dance from the working db_pairs path. This join class had no prior test coverage — the solver-level hand-calc test is its first.

Verification

• Hand-calc gate (solver): fan-at-t0 fixture with recourse + non_anticipativity_invest_periods=[period1] + non_anticipativity_periods=[] → period-1 v_invest equal across branches, period-2 per-branch, operations divergent from t0; objective RP = 173,250 at a unique LP vertex, exactly matching Slice E's hedge (shared-first-stage-by-constraint ≡ shared-trunk-by-construction), with RP − WS = 15,750 (EVPI). Divest-tie and node-side tests included.
• Full sweep: 3795 passed, 0 failed; all 4 CI schema-verify steps pass; zero golden value changes beyond the mechanical v72 diff.

Next (spec'd): construction lead time (Slice H) builds on this — per-technology lead times, where per-technology staging emerges automatically from this window once v_invest is order-indexed, plus the commissioning lag.

🤖 Generated with Claude Code

https://claude.ai/code/session_0199XGsB22JSNNHH1cxcWbcQ