§1 — The problem, with evidence

Lead, 2026-06-03: "We seem to be very good at producing new plans (the imas-efit repo has ~800) but very bad at keeping them in-sync, updating their status and archiving them once done. This needs major thought and consideration."

Measured state (reckon audit, 2026-06-03): 819 docs, 811 schema-conformant. Dashboard symptoms and their diagnosed causes:

• Shipped plans showing 0% impl. Root cause: impl is a settable scalar that the server does not auto-compute in practice (PLAN-FORMAT calls it server-computed; _set_impl/edit_plan is the only writer) — and no workflow ever set it. Fixed for the S8 dependency graph on 2026-06-03 (8 plans, commit 4f6f8402); the mechanism gap remains for the other ~810.
• Delivered RCAs sitting "active" forever. Research docs get a verdict, inform a fix, and nobody flips them to done — they then render as stale 0% dependencies on every plan that cites them.
• Collapse-on-ship (§5b) skipped under fleet pressure. The S8 orchestration resolved followups and set statuses promptly but deferred landed-summary cards and per-stage archive HTMLs "to close-out" — exactly how evergreens rot. The §5b step is mandatory in reckon-ship but nothing enforces it.
• No archive waves. plan-archived=1 exists but is rarely used; docs/completed/, docs/superseded/, docs/research/closed/ conventions are inconsistent. 819 live docs is the result.
• Plan-creation is cheap, reconciliation is manual. Every sprint mints new plans + research docs; nothing retires the superseded ones at the moment of supersession (the only moment anyone has the context).

§2 — Design principles

• Mechanisms over manual resolutions. A one-time cleanup of 819 docs without write-time invariants reverts within two sprints.
• The moment of work IS the moment of record. Status, impl, collapse, and archive transitions happen in the same session as the work that causes them (existing AGENTS.md rule — needs teeth, not restatement).
• Make the dashboard the lint. Stale state must be VISIBLE as a violation (audit ERROR/WARN), not discoverable only by the lead noticing 0%.
• Bias against new plans. A followup on an existing plan out-competes a new plan unless the work has its own lifecycle.

§3 — Proposed mechanisms (tracked below as milestones)

1. Write-time lifecycle invariants (reckon server):
   • resolve-final-followup OR status→shipped/done ⟹ require impl set in the same edit (reject or auto-set 1.0 with warning);
   • status shipped/done + no section-landed card in prose ⟹ WARN on write + audit ERROR after N days;
   • research docs: verdict-delivered marker (informs + a resolved outcome) ⟹ prompt status done.
2. Audit cadence: a scheduled (cron/CI) reckon audit + staleness sweep — plans modified >30d with status active and impl <1 get flagged into a weekly hygiene report; the lead sees a SHORT list, not 819 docs.
3. Archive waves: quarterly (or per-sprint-close) pass setting plan-archived=1 on done/superseded docs older than a threshold; consistent home (docs/archive/) for per-stage records; the dashboard default view shows ONLY live docs.
4. Plan-creation diet: reckon-create gains a pre-flight question — "does an existing plan own this lifecycle?" — and research docs default to plan-informs-linked notes rather than standalone status-bearing docs unless they carry followups.
5. Sprint-close checklist as code: reckon-edit close-sprint intent refuses to close while sprint items have unresolved followups, unset impl, or uncollapsed shipped sections — turning the S8-style deferral into a hard gate.
6. Backfill wave: one supervised fleet pass over the existing inventory (batched ~50/worker, haiku/sonnet) applying §3.1 invariants retroactively: set impl, flip delivered RCAs to done, archive superseded docs. Lead reviews the diff summary, not each doc.

§4 — Milestone progress

• M1 — write-time lifecycle invariants: blocked pending the lead's strictness decision.
• M2 — audit cadence: landed in reckon via the new reckon audit CLI command plus test coverage.
• M3 — archive waves: groundwork landed in reckon's AGENTS.md via a quarterly archive-wave policy (plan-archived=1 on done/superseded docs older than 90 days); automated sweep still pending.
• M4 — plan-creation diet: blocked pending design review.
• M5 — sprint-close gate: blocked pending the same strictness choice as M1.
• M6 — backfill wave: not implemented yet; the audit output now provides the candidate list for a later supervised pass.

§5 — Open questions for the lead

• Enforcement strictness: reject-on-write (hard) vs warn-on-write + audit-ERROR (soft) for §3.1? Hard rejection fights fleet workers; soft rot fights the lead.
• Where does the hygiene report land — reckon dashboard panel, weekly scheduled session, or CI artifact?
• Is the backfill wave (§3.6) worth ~one fleet-day now, or do invariants + organic touch suffice?

Project: imas-efit (mechanisms land in ~/Code/reckon)
Plan:    plan-lifecycle-hygiene
Section: SS3.1 + SS3.5 after SS5 answers
Tier:    opus

Context
  ~819 reckon docs; sync defects diagnosed in SS1 (impl never set, RCAs left active,
  SS5b skipped under fleet pressure, no archive waves). M2 audit command and M3 archive
  policy groundwork are landed; the strictness-sensitive gates remain open.

State to read
  GET /plan/imas-efit/plan-lifecycle-hygiene  (SS5 answers locked as decisions)
  ~/Code/reckon/PLAN-FORMAT.md + reckon/_plan_html.py + reckon/mcp.py (_set_impl, patch_plan)

Done-when
  1. SS3.1 write-time invariants landed in reckon (per locked strictness) + tests.
  2. SS3.5 close-sprint gate landed.
  3. Audit cadence (SS3.2) scheduled; first hygiene report produced.
  4. Backfill wave (SS3.6) dispatched if the lead approved it.
  5. Outcomes recorded; this followup resolved.