Memory design: the "sweep reads the raw ledger unfiltered" wording is now inexact (a non-recall filter was added)

[StarshipSuperjam]

The engine opened this item itself — you didn't create it.

What this is. A locked sentence in the memory design is now slightly inexact, and the wording owes a one-line clarification.

• What it is: a wording item about the engine's own design docs, not your project.
• Why it's here: PR #274 Slice A: skip harness-injected pseudo-turns in the consolidation sweep #359 (memory Slice A of Two uncoordinated memory stores: distilled construction lessons live in Claude Code's built-in memory, while the engine ledger is ~97% raw capture #274) makes the consolidation sweep skip the engine's own injected notifications (a <task-notification>, the /compact continuation summary) so they are not summarised as if you wrote them. The locked memory design (systems/cognitive/memory/README.md, the durability paragraph) describes that sweep as reading "the raw ledger unfiltered."
• What is actually true: the sweep now applies one narrow, non-recall filter. The guarantee that sentence protects is fully intact — every note still stays resident and recoverable, recovery still keys on the consolidation marker (never on what the sweep reads), and the separate recall-exclusion rule is untouched. Only the single word "unfiltered" is now literally imprecise. No guardrail is weakened and no behaviour the durability law promises is lost; the deliverable-gate security review judged this disclose-and-proceed, not a blocker.

What happens next. Nothing changes in the engine's behaviour; this only tracks a doc clarification you own.

• The fix is a one-line clarification in engine-planning distinguishing "the sweep reads every resident record" (the durability guarantee) from "feeds every record to the AI as fuel" (where harness scaffolding is now skipped).
• It is your call on the locked design — the build never edits engine-planning itself.
• Until then the divergence is disclosed in PR #274 Slice A: skip harness-injected pseudo-turns in the consolidation sweep #359 and here, so it is tracked rather than lost.

More detail.

• PR #359 — memory Slice A (the change that introduced the divergence)
• #274 — the memory anchor decision this slice serves