[docs] session retrospective: three meta-lessons captured

.claude/notes/automation.md

@@ -3,6 +3,26 @@
 Lessons specific to Claude Code skills, slash commands, classifier
 interactions, and parallel-agent workflows in this repo. Newest-first.
 
+### When a memory captures a recurring failure, ask whether a hook can enforce it
+
+A saved feedback memory ("do X before Y") is advisory: it relies on
+the agent discovering and applying the rule at the right moment. A
+pre-PR checklist memory landed mid-session was followed by a lint
+failure shipped to CI within the hour — the memory was in place,
+the failure mode it described still occurred. The same gap was then
+closed reliably by a `PreToolUse` hook that injects the checklist as
+a system reminder the agent cannot miss.
+
+Memory remains useful for *rationale* (six-months-cold-reader can
+look up *why* the rule exists). Hooks provide *enforcement* (the
+agent cannot proceed without acknowledgment). For any "always do X
+before Y" pattern, prefer the hook; keep the memory as the
+explanation for why the hook exists.
+
+Anchor: `.claude/settings.json` `PreToolUse.Bash` hook (PR #80
+commit `3c12d02`); related personal memory cites
+`feedback_test_plan_before_pr`.
+
 ### Match shell-command hooks by regex word-boundary, not substring
 
 Substring patterns like `*"gh pr create"*` fire on any command that

.claude/notes/review-patterns.md

@@ -4,6 +4,29 @@ Lessons about multi-lens review, adversarial passes, and iterative
 review prompt construction when running review tasks via the Agent
 tool. Newest-first.
 
+### Self-rate work, then write criteria for the next grade up
+
+Before declaring your own work done, rate it (B+, A-, A) and write
+the measurable criteria that would elevate it one letter grade.
+Implement the criteria that fit within the current PR's bounded
+scope; defer the rest to FOLLOWUPS. Re-rate. Repeat until the
+marginal cost of the next jump exceeds the marginal value.
+
+The forcing function — "what specifically would elevate this?" —
+is sharper than "anything else?" because it requires articulating
+measurable criteria. PR #76 went B+ → A → A+ in two iterations
+this way; each iteration closed real structural gaps (CI visibility
+loop, falsifying test, graduated alert tier, helper extraction,
+carve-outs, regex decomposition).
+
+Diminishing-returns inflection: A → A+ produces smaller deltas
+than B+ → A. Pay the A+ cost when explicitly asked or when the
+work is load-bearing enough to justify it; otherwise stop at A.
+
+Anchor: PR #76 commit `869450b` (final A+ state); the review-
+iteration commits leading up to it are visible in
+`git log origin/main..869450b`.
+
 ### Spawn an explicit adversarial reviewer after graded reviewers converge
 
 "Find what others missed" produces a qualitatively different signal

docs/notes/ci.md

@@ -4,6 +4,25 @@ Lessons specific to the local and remote continuous-integration
 surface. The `make ci` source-of-truth rule is in `AGENTS.md` under
 load-bearing lessons. Newest-first.
 
+### Frame CI / perf projections as ranges, not single numbers
+
+A wall-time projection labeled "X seconds" invites a "is it really
+that?" follow-up when actual diverges. Verify the projection's
+assumptions empirically before publishing, or frame as a range
+("expect 150–250s, bounded by Y") that absorbs variance the author
+has not yet measured.
+
+PR #72's verify-split was projected at 155s wall time; actual was
+242s — the projection assumed setup-go cache amortization that
+does not happen across parallel jobs (verified post-merge in PR
+#77). The cost of the unverified projection was an investigation
+round to explain the gap.
+
+Anchor: PR #72 commit `7375cef` body (the 155s claim); PR #77
+`docs/FOLLOWUPS.md` § "CI wall-time (post-verify-split)" entry
+that records the actual measurement and falsifies the cache
+assumption.
+
 ### Express CI rate-limited assertions in per-unit-of-work units
 
 CI assertions that divide a measurement by wallclock break under