# Modular Runtime Phase 4 — Long-Build Gate

Date: 2026-08-01

Status: accepted; Phase 4 gate complete

## Accepted outcome

The graph lane now carries bounded product work across multiple milestones, real model calls, compatible parallel implementation, serialized authority, deliberate controller interruption, horizon renewal, verification-cache replay, and graph compaction without duplicate work or a global reset.

Phase 4 is accepted because the long-build module now satisfies every named gate requirement:

| Required proof | Current evidence | Result |
|---|---|---|
| Architecture/interface milestone support | Version-pinned graph inputs, dependency artifacts, sealed bounded horizons, and immutable checkpoints | accepted |
| Rolling one-to-two-milestone horizon | Initial horizon plus compare-and-set exact-prefix renewal; revision 1 to 2 in the combined canary | accepted |
| Dependency-safe implementation waves | Existing graph modules start compatible cooperative work concurrently while the integration queue remains unclaimed | accepted |
| Compatible integration batches | One durable epoch verifies one deterministic union and moves authority once; post-promotion replay is exact | accepted |
| Replay-safe graph compaction | Accepted milestone detail becomes a content-addressed summary without changing readiness or deleting audit stores | accepted |
| Bounded replanning | Sol/high may renew only the exact next coarse prefix after checkpoint completion | accepted |
| Warm-cache reuse | Content-addressed trusted verification hits 5/5 exact replays and refuses stale, failed, dirty, unreaped, or corrupt evidence | accepted |
| Critical-path improvement | Final canary records 104.187 seconds of worker overlap and 1.893× worker-lane speedup | accepted |
| System-level interruption/resume | Forced post-promotion restart repeats zero model calls, reconstructs authority, and completes through a second horizon | accepted |

## Live evidence chain

| Canary | Accepted boundary | Result | Time |
|---|---|---:|---:|
| [`MODULAR_RUNTIME_PHASE4_LONG_BUILD_CANARY_01.md`](./MODULAR_RUNTIME_PHASE4_LONG_BUILD_CANARY_01.md) | Real Sol horizon, backend-to-UI dependency, running product, forced reopen | 1/1 | 1,107.973 s |
| [`MODULAR_RUNTIME_PHASE4_BATCH_CANARY_01.md`](./MODULAR_RUNTIME_PHASE4_BATCH_CANARY_01.md) | Compatible same-base workers, deterministic union, crash replay | 1/1 | 129.823 s |
| [`MODULAR_RUNTIME_PHASE4_AUTO_BATCH_CANARY_01.md`](./MODULAR_RUNTIME_PHASE4_AUTO_BATCH_CANARY_01.md) | `LongBuildModule` automatically consumes the accepted batch seam | 1/1 | 115.926 s |
| [`MODULAR_RUNTIME_PHASE4_SYSTEM_CANARY_01.md`](./MODULAR_RUNTIME_PHASE4_SYSTEM_CANARY_01.md) | Combined cache, restart, compaction, renewal, and terminal execution | 1/1 | 270.316 s |

Total execution time across the four accepted current-revision live Phase 4 canaries: 1,624.038 seconds.

Deterministic component evidence: [`MODULAR_RUNTIME_PHASE4_COMPACTION_RENEWAL_CACHE_01.md`](./MODULAR_RUNTIME_PHASE4_COMPACTION_RENEWAL_CACHE_01.md).

## Combined canary facts

- three real Luna/xhigh implementation calls;
- one real Sol/high renewal call;
- two core workers overlap for 104.187 seconds;
- one forced restart after Git authority moved but before SQLite terminalization;
- zero repeated model calls after reopen;
- three batch verification cache hits and two direct verification cache hits;
- five of five exact replay commands avoided, with evidence hashes unchanged;
- two checkpoints and two first-parent promotions;
- two immutable milestone summary records;
- one terminal active summary node at graph revision 11;
- zero human checkpoints.

## Current-revision regression

Commands:

```text
pnpm --filter @orchestrator/modular-builder-core test
pnpm --filter @orchestrator/modular-builder-core build
pnpm --filter @orchestrator/modular-builder-core check:source
git diff --check
```

Results:

- 221 deterministic tests passed across 43 active suites;
- 15 separately gated live tests were skipped in the deterministic lane;
- TypeScript passed with no diagnostics;
- 64 runtime source files average 179 logical lines, with p50 141 and p90 414;
- maximum runtime source length is 494 logical lines;
- no runtime source file crosses the 500-line review threshold;
- no live Phase 4 canary used a human checkpoint.

## Authority and mechanism review

The accepted Phase 4 implementation reuses one `WorkGraph`, one `GraphCoordinator`, one `ResourceBroker`, one integration queue, one SQLite authority, and the existing recipe modules and graph runner. It adds bounded contracts and passive artifacts around those owners. It does not add a graph database, distributed scheduler, arbitrary workflow language, nested supervisor, archive store, model-owned retry loop, or parallel integration authority.

## Exact boundary before Phase 5

Phase 4 proves long-horizon execution, not release or migration authority. Phase 5 must add the isolated `release` recipe, run matched direct and complex canaries beside legacy Core, compare autonomous contract acceptance, elapsed time, model calls/cost, retries, recovery, and later human product judgment, and switch authority only if the new runtime matches direct work and materially improves complex work. Legacy Core remains untouched and authoritative until that evidence exists.