McRitchie Studio

Generator, rake task, do-not-edit marker, and canonical-naming are all correct — the three acceptance criteria are met and CI is green. One blocking issue, both reviewers caught it: the canonical memory SOP and the GENERATED artifact assert in PRESENT TENSE that a fresh session auto-loads insights via 'bin/session-insights' (SessionStart) — but that command exists only in the sibling worktree 'feed-banked-insights-forward', not on release/main. Merged standalone, docs/agents/system/memory.md, the generated docs/agents/shared/insights.md, the MEMORY.md stub, and the doc_generator.rb header all ship to main dangling a nonexistent command and telling fresh agents 'a new agent hatches already knowing them' — so agents skip reading insights.md and the lessons are silently missed, the exact write-only-loop failure this PR exists to fix. Resolve either way: (PREFERRED, self-contained) soften all four bin/session-insights references to planned tense (e.g. 'a fresh session WILL load these via a planned SessionStart loader') so #345 never asserts a false fact regardless of merge order; OR confirm+sequence so 'feed-banked-insights-forward' merges first/with this PR and keep present tense. While in there, also align the count: lib/tasks/insights.rake prints ActionGrade.banked.count but the doc header uses post-filter rows.size — they diverge on a blank-slug banked grade; use the same source. Re-request review after.

Conductor fact-check on the block: bin/session-insights IS already on origin/release (added by #342 feed-banked-insights-forward, an assembled co-member of THIS release candidate rel-20260703-d5eb81) — verified via git ls-tree. So the reviewer's 'command exists nowhere / not on release' premise is stale, and the 'sequence the loader task to merge first' condition is already satisfied (they ride the same RC). Remaining actionable rework: (1) the SessionStart HOOK that invokes it is an install-time wiring (bin/install-agent-docs → ~/.claude), not in-repo, so soften the doc's present-tense 'a new agent hatches already knowing them' to reflect that the auto-load depends on the installed hook; (2) fix the count divergence (ActionGrade.banked.count in the rake vs post-filter rows.size in the header). Rebase on the updated release when reworking.

Resolved QA block: softened all four bin/session-insights refs to planned tense (true regardless of #342 merge order); aligned rake count with the doc's rendered count. Re-certified green.

DoR merge-gate REFUSES: full-suite + rubocop evidence is fingerprint-STALE. Recorded evidence is [full-suite@bb842f78...]/[rubocop@bb842f78...] (certified at commit cdd3eb1), but the current worktree tree fingerprints to 194d449b... — commit 475d738 'align the generated count' edited app/services/insights/doc_generator.rb AND test/services/insights/doc_generator_test.rb (plus the committed insights.md) AFTER certification. That is test-relevant source changed post-cert, so the recorded green does not cover the code being shipped, and bin/release merge would refuse it too. Code review itself is clean (both reviewers approve; render is pure/deterministic, generate! excludes unbanked, MEMORY.md retirement coherent) and the tree IS green — I ran test/services/insights/doc_generator_test.rb: 6 runs, 32 assertions, 0 failures. So this is a pure re-certification, not a real red. Fix: from the worktree .worktrees/generate-insights-doc-from-bank, run 'bin/full-suite-check generate-insights-doc-from-bank' then 'bin/dor-check generate-insights-doc-from-bank' back-to-back in ONE call (gate-same-root rule — no board writes in between or the fingerprint drifts), confirm dor-check is green, then move back to submitted.

Re-certified at current head (bb842f78822f) — stale-cert resolved; no code change beyond the approved tense/count fix.