docs(evaluator): PR-E E5 loop readiness audit

[briansrls]

Summary

• add the PR-E E5 loop readiness/blocker audit
• cross-link the E5 dispatch brief to the audit
• document LoopNode fields, Cardinality/Descent handling, frame-state interactions, and the exact resume gate before executable loop work

Scope

• docs-only readiness receipt
• no Transform/Branch/Bind implementation
• no strategy carrier widening
• no substrate, runner, or bootstrap changes

Validation

• git diff --check
• pre-push: cargo fmt --all --check

[briansrls]

Manager pass: scope looks correct for the E5 dispatch. This is docs-only, names the concrete blocker without pretending zero-iteration or fixed-body support proves loop execution, and preserves the STOP+PING boundaries: no Transform/Branch/Bind implementation, no strategy widening, no substrate edits, no runner changes.\n\nNo manager-blocking changes requested. Hold for CI / scheduled reviews.\n\n— sent from snappy-moth-795

[briansrls]

Review metadata

• Provider / model: claude / claude-opus-4-7
• Commit: 66ff1145 · Trigger: schedule
• Comparison: origin/main @ 8cd53594 ... review/pr-1401-66ff1145 @ 66ff1145
• Thinking: 17s wall

---

Docs-only PR. Cleanly scoped, faithfully describes the live substrate, and explicitly defers implementation.

Verdict: APPROVE — Docs-only readiness audit; accurately cites the live LoopNode/LoopBound shape, names the blocker (E1-only eval_node coverage), keeps scope to a receipt, and lists explicit non-goals. No invariant or modeling-discipline concerns in the diff.

[briansrls]

Review metadata

• Provider / model: codex / unknown
• Commit: 66ff1145 · Trigger: schedule
• Thinking: 193s wall

---

BLOCKING (1)

Root Cause

• docs/briefs/r3-pr-e5-loop-readiness-audit.md Loop count decoding is specified as Int-shaped value decoding instead of a Nat/cardinality witness contract → require non-negative finite count decoding and add a negative-count fail-closed test before E5 can implement from this receipt.

⚠️ One blocking readiness gap remains in the future E5 acceptance contract.

[briansrls]

BLOCKING: The rule allows a runtime integer literal count but never requires negative Int values to fail closed, so E5 could accept a non-cardinality loop bound despite the Decidability invariant's bounded-iteration contract.

[briansrls]

Review metadata

• Provider / model: cursor / composer-2
• Commit: ac8710fa · Trigger: schedule
• Comparison: origin/main @ 8895a5e4 ... review/pr-1401-ac8710fa @ ac8710fa
• Thinking: 23s wall

---

Findings: None. The change is documentation only: a new readiness brief plus a short cross-link in r3-evaluator-dispatch.md. Nothing here conflicts with INVARIANTS / modeling-discipline (it explicitly records fail-closed behavior, defers Descent, rejects parallel loop carriers, and names a resume gate and test bar). CODING.md and TESTING.md govern Rust/tests; nothing in this diff touches them.

Verdict: APPROVE — The diff is small, honest about scope (“docs-only receipt”), and documents the E5 blocker and acceptance tests without landing substrate or implementation debt. No violations of the referenced rubric grounded in these edits.

[briansrls]

Review metadata

• Provider / model: openai-pro / gpt-5-5-pro
• Commit: a8262c71 · Trigger: manual
• Comparison: main @ 8895a5e4 ... feat/r3-evaluator-e5-loop-readiness @ a8262c71
• Conversation: View conversation

---

1. Story of the diff

This PR turns E5 loop execution from a vague future acceptance target into an explicit readiness audit. docs/briefs/r3-evaluator-dispatch.md:107-110 now points readers to the new audit and states the core blocker: E5 cannot honestly prove accumulator threading until the body evaluator can run a loop body that consumes the iteration accumulator binding. The new audit documents the live LoopNode / LoopBound shape, defines the intended cardinality-loop execution rule, requires fail-closed count decoding, leaves descent loops as a named residual, and sets a resume gate plus required tests for the later implementation PR. The main mechanism is deliberately not implementation: docs/briefs/r3-pr-e5-loop-readiness-audit.md:3-5 says this is a docs-only receipt and does not implement eval_loop, change substrate carriers, widen strategy carriers, or add runner behavior.

2. Invariant categories

1. LAYER MODEL — Compliant. The diff discusses substrate-facing loop carriers but does not modify substrate; it explicitly says E5 should consume the existing LoopNode / LoopBound shape directly and “must not add a parallel loop carrier or reinterpret LoopBound with string labels” at docs/briefs/r3-pr-e5-loop-readiness-audit.md:34-37, with substrate edits also excluded in the non-goals at docs/briefs/r3-pr-e5-loop-readiness-audit.md:124-128.
2. INVARIANTS.md + modeling-discipline.md — Compliant. Fail-closed is handled directly: count decoding must accept only non-negative runtime integer literal counts, while missing, non-integer, or negative counts must produce typed diagnostics rather than defaulting to zero or wrapping unsigned at docs/briefs/r3-pr-e5-loop-readiness-audit.md:57-60. Single-authority / facts-flow-forward is also preserved by naming LoopNode.source as the candidate accumulator binding authority while requiring STOP+PING if lowering says another port owns it, rather than inventing a second authority, at docs/briefs/r3-pr-e5-loop-readiness-audit.md:81-87.
3. CODING.md — Compliant. Although this is docs-only, the proposed implementation shape follows the repo’s data + free-functions style: eval_loop is specified as a free function over explicit inputs &Dag, &LoopNode, &mut EvalStateStack<Value>, and &EvalStrategy, returning Result<Value, EvalError> at docs/briefs/r3-pr-e5-loop-readiness-audit.md:44-50.
4. TESTING.md — Compliant. No tests are required in this PR because no executable behavior changes, and the audit avoids dead API. The future implementation is constrained with behavior-driven required tests covering zero iteration, multiple iteration, accumulator visibility/threading, missing/non-integer/negative counts, descent residual, and stack restoration at docs/briefs/r3-pr-e5-loop-readiness-audit.md:107-118.
5. LOCKED DESIGN DECISIONS — N/A. The diff references parent dispatch/scaffold authorities at docs/briefs/r3-pr-e5-loop-readiness-audit.md:7-9, but it does not reference or alter anything marked locked.
6. TRACKED vs UNTRACKED DEBT — Compliant. The new blocker is tracked rather than left as ambient TODO debt: it is documented as an implementation blocker at docs/briefs/r3-pr-e5-loop-readiness-audit.md:3-5, bounded by scope/non-goals at docs/briefs/r3-pr-e5-loop-readiness-audit.md:120-128, and given a concrete dissolution/resume trigger at docs/briefs/r3-pr-e5-loop-readiness-audit.md:99-105.

3. Verdict

APPROVE. The PR is a docs-only readiness receipt that makes the E5 blocker explicit, preserves existing substrate authority, avoids fake partial implementation, and gives the later implementation PR a bounded resume gate and test checklist. I found no diff-anchored issue requiring changes.

[briansrls]

Review metadata

• Provider / model: claude / claude-opus-4-7
• Commit: a8262c71 · Trigger: schedule
• Comparison: origin/main @ 8895a5e4 ... review/pr-1401-a8262c71 @ a8262c71
• Thinking: 11s wall

---

APPROVE — Docs-only PR adding a readiness audit for E5 plus a cross-link from the parent dispatch brief. Scope is clearly bounded ("does not implement…", explicit Non-Goals), the loop shape it describes matches the substrate it cites (LoopNode/LoopBound), and it correctly defers Descent as a named fail-closed residual rather than papering over it. No invariant, modeling, coding, or testing rules are touched by the diff.

[briansrls]

Verified the earlier codex blocker against current head a8262c71: resolved. The readiness audit now requires count decoding to accept only a non-negative runtime integer literal count, and explicitly says missing, non-integer, or negative counts must fail closed rather than defaulting to zero or wrapping. The future implementation test list also includes LoopBound::Cardinality with negative integer count fails closed.\n\nCurrent checks: fmt, ci, and v3 are passing; self_host_ratchet is still pending. No manager patch needed.\n\n— sent from snappy-moth-795