[jussi-modernpath]

Why

Cold review on recent epics ran four to ten rounds (EPIC-ANALYSIS-009: 10, EPIC-CLI-022: 9, EPIC-OBAN-001: 8) although PROCESS.md bounds it at two. The rounds were not wasteful: each one found real problems in the packet. Two causes, from the finding history and the EPIC-ANALYSIS-009 trial on 2026-09-25/26:

1. The packet never stated the state model the change touches — writers, readers that branch on it, crash mid-write, staleness, recovery — so every round re-derived it and found a different gap.
2. Findings were resolved by disposition edits that added mechanisms (a retry, a guard, a task link) the packet had never reconnoitred; the next round faulted those. Rounds 3, 4 and 6 on EPIC-ANALYSIS-009 found defects only in mechanisms the previous round's resolutions introduced.

With a state inventory in the packet, rounds 5 to 8 found nothing new about the stated state and only faulted late mechanisms or a row the inventory lacked. The reviewer of round 5 wrote that the inventory turned the review into a diff against a stated set.

The remedy therefore makes the packet complete before review rather than limiting rounds.

What changes

• PROCESS.md §Entry packet, item 3 — reconnaissance carries a state inventory whenever the change adds or touches persisted or shared state; every mitigation cell ends closed, residual or decided.
• PROCESS.md §Planning and readiness — a packet is complete before its review starts: a row the inventory lacks is a finding, a cell closed with "backlog" while a criterion asserts it is a finding, a mechanism the packet does not name is not part of the change. A RESOLVED disposition names its kind (packet edit, scope action with its record, decision with its USER: source). No new mechanism enters a packet during a review cycle; a finding whose fix needs one is a scope question. A round whose material findings fall on mechanisms an earlier resolution introduced is evidence the packet was reviewed incomplete.
• rdd-plan step 4 — how to write the inventory (join rows as much as entities) and the self-audit against it before handoff. A value the packet states is read at the call site that applies it, not at its definition — taken from PR 22 (@paveq), the one hunk there that raises packet completeness rather than managing rounds. Step 7 — a mechanism a review needs later goes back through scope and reconnaissance, never into the packet.
• rdd-cold-review step 2 — enumerate the inventory's cells and the transitions between concurrent writers in both directions; a missing inventory on a state-touching change is the first material finding. Step 6 — the three resolution kinds; a resolution that adds a criterion, widens a boundary or adds a hop is a scope question returned to rdd-plan. Report — name the findings that fall on late mechanisms.

Not changed

The two-round paragraph in PROCESS.md and rdd-cold-review step 8 stay as they are; this PR adds no caps. rdd-deliver, rdd-build, rdd-start and file-state/ are untouched.

Relation to earlier PRs

Supersedes #13 (closed 2026-09-26) and #22. From #13 it keeps the idea that a finding introduced by a remediation is a distinct fact, as prose in the report rather than a classification enum, and leaves out the plan-subject fingerprint, coverage rubric, continuation gates and CRF- records. From #22 it keeps the call-site rule only.

The CLI half — a required resolution kind on process findings disposition, a per-round widening count in process findings list, and a plan-phase check for the inventory section — is a separate epic in modernpath-v1 (BACKLOG-TOOL-218).

Validation

• git diff --check
• node --check skills/rdd-audit/audit-citations.mjs
• node skills/rdd-audit/audit-citations.mjs PROCESS.md file-state README.md (no citations in these documents, as before)

🤖 Generated with Claude Code

2026-09-27 — after the first trial (modernpath-v1 EPIC-CLI-027, two cold-review rounds): four rules added. Reconnaissance starts from the ledger (the delivered requirement that owns each touched surface, its acceptance carried as a constraint); generated system documentation is read before the code and cited; an inventory row lists each write shape it admits; a resolution edit re-enters reconnaissance for what it names and a reviewer's proposed fix is a claim. Each would have caught one of the round-2 findings.

2026-09-28 — after EPIC-CLI-027 was built and its pull request reviewed: two rules added. A review of the built change is not a round of the packet's review: it runs from its own independent context and numbers its own rounds, outside the packet's round history and the two-round bound, and an open material finding on it stops delivery (rdd-completion-review step 2 now names it). A RESOLVED finding on the change names the commit that resolved it, a fourth kind beside the three, and the commit only brings the code to what the approved change already contained. A migration or a schema change is production and lands with the GREEN it serves (PROCESS.md RED paragraph, rdd-plan step 5). The CLI half — the commit kind and a review context for built work — is filed in modernpath-v1 as BACKLOG-TOOL-274 and BACKLOG-TOOL-279.