Improve iterative design-validation loop for highly technical/impactful changes

[JRedeker]

Context

During the improveCrossProject change (a highly technical, high-impact change touching cross-project durable state, Temporal workflow semantics, and multi-store consistency), the design phase required 6 iterations of independent validator (adv-researcher) review before reaching convergence. Each iteration surfaced 2-6 genuine architectural findings.

Iteration finding-count trend: 5 → 6 → 4 → 3 → 4 → 2 (slowly converging).

Problem

For highly technical / impactful changes, the current design-validation loop has these friction points:

1. No upfront classification of "high-technical-impact" — the loop topology (single validator pass → approve) is identical for trivial and structural changes. High-impact changes need more rigorous upfront scoping.
2. Validator finds new concerns each round — each iteration is genuinely valuable but expensive in agent context. No mechanism to bound iteration count or escalate to human architect.
3. No "design debt" surface — when iterations hit practical limits, there is no first-class way to defer findings as fast-follow changes while proceeding with documented design debt.
4. No multi-validator triangulation — single adv-researcher pass has one perspective. High-impact changes might benefit from adv-tron (structural) + adv-researcher (external) + adv-reviewer (implementation feasibility) cross-check.
5. Agreement artifact evolution is implicit — the agreement changed mid-session (added AC11-AC15 + avoidance 6), invalidating earlier design assumptions. No versioning signal to the design phase.

Proposal

Add an "impact tier" classification at proposal/agreement time that gates design-phase rigor:

Tier	Trigger	Design-phase path
standard	Single-component, well-understood surface	Single validator pass; approve on PASS/CAUTION
high-impact	Cross-cutting, durable-state, authn/authz, workflow semantics, multi-store consistency	Multi-validator triangulation + bounded iteration budget (default 3) + design-debt escape hatch
critical	Security-bearing, data-loss-bearing, core-contract changes	Above + mandatory human architect review at design gate

Specific additions:

• Impact classification field in proposal artifact (impact_tier: standard | high-impact | critical); machine-enforced; human override allowed.
• Bounded iteration budget — after N validator rounds (default 3 for high-impact, 5 for critical), if still CONFLICT, auto-pause and offer: continue / accept-with-debt / descope / human-architect.
• Design debt first-class — adv_change_update_design_debt tool that records deferred findings as fast-follow ADV child changes linked from parent.
• Multi-validator triangulation — adv_design_validate_multi that runs adv-researcher + adv-tron + adv-reviewer in parallel and synthesizes.
• Agreement versioning — agreement artifact carries revision field; design artifacts reference agreement_revision; mismatch triggers re-validation.

Evidence

Observed directly on improveCrossProject (2026-07-20):

• 6 validator iterations required
• Each iteration consumed meaningful agent context (design body grew from ~2k words to ~5k words)
• Several findings were genuinely load-bearing (Signal-cannot-return, cross-store atomic rollback impossible, CWE-367 race) and would have shipped structural debt if skipped
• No mechanism to bound iteration or escalate short of user-initiated pause
• Agreement artifact was updated mid-session; design assumed stale snapshot for one round

Out of scope

• Replacing the validator (adv-researcher works well; this is about loop topology)
• Changing the gate sequence (proposal/discovery/design/planning/etc unchanged)
• Auto-approving designs (always human checkpoint at design gate)

Related

• Change: improveCrossProject
• Iteration logs: improveCrossProject|change:researcher:design-validation-{v1..v6}|adv-researcher|{1..6}