P1+P2: runnable engine + artifact lock, schedules/drivers/fallback audit (#23–#26)

[ebootheee]

Closes #23, #24, #25, #26. Two commits on this branch.

P1 (#23 + #24) — reliably emit a runnable engine.js

A clean ete init on the real models never finished: the chunked emitter built the cell-level dependency graph as a full in-memory map + a second serialized String (OOM on multi-million-cell models), and engine.js was emitted after it, so the runnable engine never landed.

• Emit engine.js before the dependency graph (it depends only on the sheet DAG + partitions); stream dependency-graph.json to disk one entry at a time (the OOM fix). Output stays deterministic; schema unchanged.
• ete init --timeout <seconds> (default 1800; 0 disables); verify engine.js after a fresh parse (fail loud, never partial; --reuse-parse exempt).
• chunked/build-manifest.json (Artifact must be consumable downstream without per-version reconciliation (lock layout + emit content hash + always emit engine.js) #24): locked artifact layout + a stable contentHash over the identity artifacts (engine.js, sheets/, _ground-truth.json, manifest.json) — stable across rebuilds, changes on drift. New lib/build-manifest.mjs, npm run test:runnable + CI step.

P2 (#25 + #26) — contract schedules, drivable inputs, fallback audit

• Schedules/distributions (named-outputs must support pinning value-bearing cells (per-class MIP Proceeds, hurdle, valuation) — not just MOIC/IRR #25): per-year time-series outputs (distributionsToEquity, outstandingDebt, equityBase, freeCashFlow, per-class classes[].distributions) pinned into named-outputs.json as cellRange + perYear[], brought under the dependency closures via expandRange().
• Drivable inputs (named-outputs must support pinning value-bearing cells (per-class MIP Proceeds, hurdle, valuation) — not just MOIC/IRR #25): exitMultiple, exitYearSelector, hurdleRate pinned into named-inputs.json.
• Fallback audit (Emit _fn-fallbacks.json — audit of cells using unsupported-function fallbacks (correctness gate) #26): scan generated sheet modules → _fn-fallbacks.json; flag any named output/schedule whose closure passes through an _fn() stub. Reports by default (annotates outputs with resolvesThroughFallback + stats.fallbackViolations; ete init warns); hard-fails under --assert-no-fallbacks (CI / golden-master gate) so it doesn't block the real models (~11,813 fallbacks today).

Review-hardening

The #26 gate as first written throwew mid-emit and ete init's try/catch swallowed it — silently dropping the entire contract (named-outputs/inputs/cell-types) while still reporting success, a failure that would fire routinely on the real models. It now emits all maps first, then surfaces the result so it can never be swallowed (reproduced + verified end-to-end).

Tests

Full matrix green: npm test 387 (test-manifest-maps 47→57), smoke 78/78, test:runnable 20, test:depgraph 11, test:engine 21, test:slimming 13, cargo test 11/11 (0 warnings). CI runs ubuntu + windows.

Known follow-ups

• Drivers under --reuse-parse without the workbook are skipped (work in the normal init path).
• Schedule scalar baseCaseValue sums across years (right for flows, not balances; perYear is authoritative).

🤖 Generated with Claude Code