audit: canonical-type-names milestone close (clean) + ratify post-23.3.3 compile baseline

docs/JOURNAL.md

@@ -14506,3 +14506,69 @@ Workspace at iter-ct.4 close: full `cargo build --workspace`
 and `cargo test --workspace` green (457 tests, 0 failed — net
 +6 over ct.3's 451: 2 pin tests in hash.rs, 2 prose trim unit
 tests, 1 prose snapshot, 1 E2E test).
+
+## 2026-05-11 — Audit close: canonical-type-names milestone
+
+Audit pipeline ran clean per skill protocol.
+
+**Architect drift review:** zero items. DESIGN.md's Decision 2
+amendment matches what the code enforces; the four obsolete
+mechanisms in the ct.1 catalogue are all gone or correctly
+narrowed; `qualify_local_types` is applied uniformly at all three
+cross-module fn-lookup sites; the five remaining "bit-identical"
+paragraphs in DESIGN.md refer to features orthogonal to the
+canonical-form migration and remain factually correct. JOURNAL
+test-count claim (457 at iter close) verified against
+`cargo test --workspace`. Spec acceptance criteria 1-8 empirically
+satisfied.
+
+**Bench-regression check:** `bench/check.py` and `bench/cross_lang.py`
+green at exit 0 (0 / 88 metrics regressed). `bench/compile_check.py`
+initially reported 23 / 24 regressed (compile-time +27-44%
+across every fixture). Bencher dispatched to localise; the
+result refuted the initial hypothesis (validator-causes-regression)
+with N=15 median measurements: validator's measured CPU
+contribution is 0.01-0.07 ms per `ail check`, within noise.
+
+Localised to **iter 23.3.3** (`12d9130` — `Ord extends Eq` + 3
+Ord instances added to `examples/prelude.ail.json`). The prelude
+grew from 18 to 223 lines between `3742583` and `12d9130`. The
+cost is paid on every `load_workspace`: parse, canonicalise,
+typecheck the prelude module, register its 6 instances. The
+bench fixtures are tiny (hello = 4 defs); the prelude is now
+larger than every user fixture in the corpus and dominates load
+time. ct.1.* contributes a residual +0.05-0.15 ms on top — not
+material against the ~+20 ms build floor.
+
+**Resolution:** ratify. The prelude growth is a deliberate
+post-iter-23.3 design state (Eq + Ord classes with three primitive
+instances each are the canonical typeclass surface; the prelude
+is auto-loaded for every workspace per iter 23.1.2). Constant
++17 ms per `ail build` invocation is the documented price of
+the canonical typeclass prelude. Baseline rebased via
+`bench/compile_check.py --update-baseline`; post-rebaseline run
+is 24 / 24 stable.
+
+Lesson logged: when a milestone closes and a bench script reports
+broad regressions, the canonical move is to localise BEFORE
+ratifying. Without the bisect, ratifying would have written
+"canonical-type-names caused +30% compile time" into the
+JOURNAL — a false attribution that would mislead future
+archaeology.
+
+**Rustdoc:** 20 pre-existing-pattern warnings (18 carried from
+earlier, +2 introduced by ct.1.5a's `type_def_module` doc-link
+and ct.3.2's `apply_per_module_types_overlay` rename). The
+existing P2 roadmap entry "Rustdoc warning sweep" covers them;
+not promoted to a tidy iteration on this audit.
+
+**Audit verdict:** carry-on. Canonical-type-names milestone is
+closed. Skill pipeline next step: `fieldtest`.
+
+Follow-up not from this audit but worth flagging as a P3 idea:
+optimising the prelude load path. A `OnceCell`-cached parsed
+prelude (or a content-addressed disk cache) would recover most
+of the ~17 ms per-invocation cost. Quantified by the bencher
+above as worth doing only if the prelude-loading cost is judged
+structurally permanent. Add to roadmap as P3 if the next
+fieldtest surfaces it as friction.