eval(deltawire): integrate treatment-fidelity preflight

[bigboateng]

Why

DeltaWire's benchmark preflight needed deterministic task contracts, binary parity, independent semantic verification, and fail-closed treatment attribution before any larger experiment could be considered.

This clean integration PR projects the reusable implementation from the conflicted experiment lineage onto current main. PR #99 remains the provenance record for raw experimental evidence.

What changed

• Adds the DeltaWire runtime hardening, deterministic plan/schema handling, contract-driven task projections and oracles, 12 canonical fixtures, and 96 required mutation failures.
• Adds frozen manifests and matrix validation, Harbor 0.20.0 gates, identical Linux/amd64 binary staging, structured Gemini shell/lifecycle parsing, independent plan-contract checks, and fail-closed B0 contamination checks.
• Adds the versioned v2-v4 probe infrastructure and the v4 environment/treatment/end-to-end receipt boundaries.
• Adds compact readiness records, receipts, reports, and an explicit integration-projection record bound to evidence commit c200dc0eb3c4dff0e4732cc4b2daa1cd535703e3.
• Excludes raw trajectories, Harbor logs, expanded generated output, and other large run evidence from this branch; those remain on eval/deltawire-preflight-v4 and PR eval(deltawire): prepare 72-run preflight and semantic oracles #99.

Observed v4 outcome

The one authorized D1-only v4 probe passed semantic exact match and treatment-use fidelity. Harbor's verifier generated a passing environment receipt, but the receipt was not retained from the main task container. The fail-closed end-to-end status is therefore fail; no probe-release.v2 was created and the range-large/r1 pair was not launched.

READY_FOR_72=false. No savings, cost, or benchmark claim is made.

Validation

• bash labs/20-deltawire/scripts/validate.sh
• Go tests, race tests, 20x repeat tests, vet, runtime-boundary and determinism checks
• 12 canonical task contracts and 96 mutation failures
• Plan-contract acceptance/rejection, structured fidelity, rerun/release/full-mode, matrix, binary-parity, readiness, and production-cleanliness checks
• No live model calls in CI

[github-actions]

Pitot coding-agent E2E

Intelligence Flow is the verification source; Pitot's public README carries the latest main status.

Agent	Ubuntu	macOS	Windows	Result	Evidence
Claude	✅ Pass · real CLI	✅ Pass · real CLI	✅ Pass · real CLI	✅ Passing	All platforms are required
Cursor	✅ Pass · hook verified	✅ Pass · hook verified	✅ Pass · hook verified	✅ Passing	All platforms are required
Codex	✅ Pass · hook verified	✅ Pass · hook verified	✅ Pass · hook verified	✅ Passing	All platforms are required
Gemini	✅ Pass · hook verified	✅ Pass · hook verified	✅ Pass · hook verified	✅ Passing	All platforms are required
OpenCode	✅ Pass · hook verified	✅ Pass · hook verified	✅ Pass · hook verified	✅ Passing	All platforms are required

Source commit: 5add6e4c5946