session 2026-06-08: ARC W3R Stage 2 (@layer middle path, !important 56→12 live) — session memory + ledger + Fowler skill-harvest + verification-ladder (delayer-diff + token-strip/render-proof cautions)

claude/memory/MEMORY.md

@@ -36,6 +36,9 @@ permalink: claude-memory/memory
 - [L1 reliability](project-L1-reliability.md) — canonical workstream tracker established 2026-05-28 (Symmetria-pulse decision). Current state at top (updated 2026-06-06) + chronological log 2026-03-21→. **Status: the 2026-06-06 three-note reply arc is SENT (`b255e74`/`0ecb45e`/`2e87bab`) — baton with Seb; do not re-enter L1 until he responds/pushes.** mindfabric-00 root-caused (temporal chain path) + restart-as-holding-pattern + backed up; PENDING-27 awaits jurist. Read tracker at /wake-up before composing L1 portion of any briefing.
 
 ## Active Session
+- [Session 2026-06-08 — ARC W3R Stage 2 (@layer), the middle path, !important 56→12 shipped live](session-2026-06-08-arc-w3r-stage2-layer-middle-path.md) — **ARC Wave 3R Stage 2 COMPLETE + PUSHED + DEPLOYED + LIVE-VERIFIED** (`5b93f0a`→`40629e8`; animalrationiscapax.org: @layer present, 12 !important, 34139 B). Read Fowler chs 1–3 first (two hats = byte-gate; Bad Smells = census lens). Census: 56 live (raw 70 = 56 + 10 dead archive-v1 + 4 comment mentions; measure-don't-speculate). **THE DESIGN — the middle path**, ruled through the Fowler/Alexander lens (Fowler IS Alexander: Cunningham/Beck→Kerievsky): a 5-layer first attempt PROVED byte-identical-after-delayering but the steward's render-review caught **3 inversions** (@layer overrides specificity; role≠file) → REVERTED → `@layer base, overrides` (Option 1, Yagni). North-star **3rd amendment "the language and the metre"** (`fc24cbe`): twelve scales = language of the TREE (legibility by composition, Stage 3), @layer = metre on the HARD LINES (enforcement where it does cascade work; promotion test). Execution all rendered-identical, steward-confirmed each step: 2a `5b93f0a` (@layer + future-rule placement rule in _foundation; _frontispiece split in-place), 2b `dd6c9d1` (34 compass !importants→win by layer), mobile-blanket retire `6579f93` (targeted nothing), glimpse close `772e0d5` (`.glimpse img`→scoped overrides; 21→**12 a11y floor**). archive-v1 KEPT (governed archive, not dead code — corrected own framing). **Kerievsky→Chamber HELD** (background agent): PDF is degraded IA scan, code listings garbled; pipeline's state-integrity refused to ship; trail → chamber-library `a01bc69` (unpushed); steward to source clean EPUB. **Pulling thread: W3R.2 Stage 3 — the spec-true tree (reorganize partials largest→smallest per the twelve scales; reify a new @layer only where it passes the promotion test); opens by reading Kerievsky framing chapters (clean in the scan).** Literal Q: reading main.scss top-to-bottom vs the spec section structure — which partials already sit right, where does the spec want the rest, is @use order even the right starting frame? Sub: Kerievsky EPUB sourced? Seb moved (be/L1)?
+
+## Archived (2026-06-07 evening — be + memory-verdicts)
 - [Session 2026-06-07 (evening) — be §2 + memory-verdicts night](session-2026-06-07-evening-be-entry-memory-verdicts.md) — BE discharged: LAUNCH+pattern-language read, note to Seb SENT (post-wrap), **a11y-gate branch BUILT + PUSHED** (axe 0 ×47 pages, Lighthouse 100s); ollama healed (broken 0.30.6 bottle → `ollama-app` cask; mindfabric 829-conn leak → A-class L1 finding, note-don't-send; login-item added); pre-commit /dev/tty fixed `dff8b7c` (April precedent `dcbba60` found late — grep-memory-before-infra-fix); **Fowler → Chamber** `a0c5991` (craft_of_software/); **Basic Memory RETIRED** (false-clean sync = disqualifier; criteria honored); **MemPalace ruled witness-not-notary + steward AMENDED CLAUDE.md** (conflict = verification trigger); lessons 7–13 → CM-AI `6cb3fe9` (**§13: every system treats the index as the memory + health as flag-not-measurement**); 3.4.0 upgrade plan saved (backup-first; migrate-wings/repair EXCLUDED) + MEMORY.md-archive + wake-canary in same slot. **Pulling thread: ARC Wave 3R tomorrow — READ Refactoring chs 1–3 FIRST (two hats; Bad Smells), then `!important` re-census, then @layer Stage 2.** Literal Q: is the census still 56 post-Waves-2-3 and what does each survivor defend? Sub: Seb replied / reviewed the branch? Seb moved (be/L1)?
 
 ## Archived (2026-06-07 afternoon — Waves 2+3 + rulings)

claude/memory/reference-verification-ladder.md

@@ -19,6 +19,10 @@ Proven gates, each earned from a real catch. Reach for the one the claim's shape
 - **Byte-identical compile gate** — prove with `cmp`/`diff` at BOTH the direct-invocation level and the full-pipeline level. Proven 3×: SCSS Stage 1 (`c37a3bd`), site.hs listing generator, §I.k transform dormancy.
 - **Expected-delta gate** — when diff=0 is impossible, prove the delta's *complete shape*: `sort -u` over all per-file diffs; every line must belong to a named class (2026-06-04).
 - **Delta classification** — corpus-wide variant: classify EVERY changed line into named classes; the unclassified residue must be exactly empty-or-whitespace. Catches what sampling misses, including one's own shipping comments (decompose + SEO passes, 2026-06-05).
+- **Delayer-diff** (2026-06-08) — for an `@layer`/structural-wrap refactor where the compiled output gains uniform wrappers: strip the `@layer X{`/closing-`}` wrappers from the compiled CSS (brace-aware script), diff vs the true baseline; byte-identical proves nothing *moved, dropped, or changed* — only got wrapped. Proven 3× in W3R Stage 2a. **CAUTION — byte-identical ≠ rendered-identical under `@layer`:** `@layer` overrides specificity, so a structural-wrap byte-proof is *necessary but not sufficient*; render-review is the gate when the cascade *basis* changes (a 5-layer wrap was byte-true yet produced 3 render inversions the diff could not see — the steward's eye caught them). Rendered-identity is then argued *by construction* (only already-winning rules promoted) AND render-reviewed — not inferred from the byte-diff.
+
+## Mechanical text transforms ("strip/replace token X across the file")
+- **Token-strip hits prose** (2026-06-08) — a whole-file `sed s/…!important…//` (or any token strip) also removes the token from comments, code-refs, and docstrings, silently mangling prose. Scope the transform to declaration lines, OR compare raw-before vs clean-declaration counts and verify comment integrity after (caught 4 mangled comments via the 56-vs-70 count mismatch; restored). Same family as *Count first, then look* — the word lives in prose too.
 
 ## Coverage claims ("all X were handled")
 - **Exhaustive route/rule census** — enumerate ALL routes/rules/sections before classifying any; classify each explicitly; no sampling. (The censused-routes discipline; §I.k.c measuring read.)

claude/memory/session-2026-06-08-arc-w3r-stage2-layer-middle-path.md

@@ -0,0 +1,180 @@
+---
+name: session-2026-06-08-arc-w3r-stage-2-layer-the-middle-path-important-56-12-shipped-live
+description: "ARC Wave 3R Stage 2 complete and live — cascade declared as @layer base, overrides; every reducible !important retired (56→12, the a11y floor). The design was ruled through the Fowler/Alexander lens (the middle path: twelve scales = the language of the TREE, @layer = the metre on the HARD LINES); north-star 3rd amendment records it. Pulling thread: W3R.2 Stage 3 — the spec-true tree — opening by reading Kerievsky's framing chapters."
+metadata: 
+  node_type: memory
+  type: project
+  originSessionId: 69ced5cd-4504-4a05-ba7a-357c77529189
+---
+
+# Session 2026-06-08 — ARC Wave 3R Stage 2: the @layer middle path
+
+Woke ~10h after the 06-07 evening wrap. Pulling thread confirmed: ARC Wave 3R SCSS day,
+opening by READING Fowler's *Refactoring* chs 1–3 in the Chamber before touching code. It
+held; the whole day flowed from it. Symmetria active throughout (ledger
+`session-ledger-2026-06-08.md`).
+
+## Past — what we did
+
+### 1. Read Fowler, Refactoring chs 1–3 (Chamber, craft_of_software/)
+Ch1 = the rhythm (compile-test-commit after every tiny step; revert-and-redo-smaller on a
+failed gate). Ch2 = the two hats (refactor vs feature, never both at once — *this is our
+byte-identical gate as stated discipline*) + Yagni + measure-don't-speculate (the date-range
+sidebar). Ch3 = Bad Smells, the lens for the !important census (Mutable/Global Data,
+Duplicated Code, Shotgun Surgery, Speculative Generality, Comments-as-deodorant).
+
+### 2. The !important census (the literal question, answered by MEASURING)
+Raw `grep -c '!important'` = 70, but **clean live declarations = 56** (the audit number
+holds): 70 = 56 live + 10 dead `archive-v1` (unimported) + 4 comment mentions of the word.
+Measure-don't-speculate caught two would-be misclassifications: the visually-hidden mixin's 9
+are LIVE (`@include`d at compass:97), and `_base` has 3 real decls not 4. Classification by
+Fowler ch3: **44 wrong-layer** (34 compass-point fighting `_typography` `a` links + 9 glimpse
+full-bleed + 1 mobile-blanket) · **12 genuine a11y** (3 reduced-motion + 9 visually-hidden) ·
+10 dead.
+
+### 3. THE DESIGN — the middle path (the day's intellectual core)
+- First attempt: a 5-layer role taxonomy (reset/base/layout/components/utilities), whole-file
+  wrap. Byte-identical-after-delayering PROVEN — but the steward's render-review caught **three
+  rendering regressions** (mobile epigraph left-aligned, `.fragment` narrower, epigraph/attr
+  spacing). Root cause: **@layer overrides specificity**, so wrapping a heterogeneous partial
+  promotes its low-specificity blanket rules above components that beat them by specificity.
+  Lesson: **role ≠ file**; my byte-diff proved CSS-text identity but NOT rendered identity —
+  the honest-degradation gap I'd flagged; the steward's eye was the real gate. **Reverted.**
+- Steward's question: *"which would Fowler do?"* → **Option 1: `@layer base, overrides`** (Yagni
+  over speculative generality; the 5-layer/12-layer taxonomy is the imposed master plan both
+  Alexander and Fowler reject). base = the whole sheet (cascade = baseline exactly); overrides
+  = only rules that must beat base by force.
+- Then: *"does the north-star make sense through the Fowler/Alexander lens — the middle path?"*
+  (Fowler IS Alexander: Cunningham/Beck carried *A Pattern Language* into software; Kerievsky's
+  *Refactoring to Patterns* is its descendant.) Verdict: the **dream** is right and is the
+  Timeless Way; the **2026-06-05 amendment over-loaded @layer** with the whole dream — it
+  conflates *legibility* (a composition property → the file tree) with *enforcement* (@layer),
+  declares the whole taxonomy up front, and imposes layer-precedence over specificity
+  everywhere (proven harmful by the 3 regressions). **The middle path:** the twelve scales are
+  the **language of the TREE** (legibility by composition — Stage 3's file organization);
+  @layer is the **metre on the HARD LINES** (enforcement only where it does cascade work).
+  Promotion test: *does a layer do cascade work, or is it pure label?* → **north-star 3rd
+  amendment "the language and the metre"** (committed `fc24cbe`).
+
+### 4. Execution — W3R.1 Stage 2 COMPLETE, shipped, live (56→12)
+All rendered-identical, steward-confirmed at EVERY step:
+- **2a** `5b93f0a` — `@layer base, overrides` declared in _foundation (first emitter,
+  authoritative) + the future-rule placement rule written down there; 18 partials whole-wrapped
+  → base, _compass whole → overrides, _frontispiece split in-place (epigraph→base,
+  .frontispiece-compass→overrides), _variables/_mixins unwrapped. Verified: delayer-diff
+  byte-identical (34273 B) + rendered-identical by construction.
+- **2b** `dd6c9d1` — removed the 34 compass-point !importants (win by layer now). 56→22.
+  reduced-motion keeps its own !important so still-stills-on-RM preserved.
+- mobile-blanket retire `6579f93` — `* { animation-duration:.2s !important }` targeted nothing
+  (zero @keyframes in build), wasn't the a11y mechanism, was a latent vignette landmine.
+  Speculative Generality. 22→21.
+- glimpse close `772e0d5` — `.glimpse img` (9 full-bleed !importants) EXTRACTED from the mixed
+  `.glimpse` block into a scoped `@layer overrides { .glimpse img }` (no !important). It must
+  beat several base img rules across DOM nesting (`.page-content > img` 0,1,1 tie; `> p > img`
+  0,1,2) — genuine cascade work → earns the overrides promotion. 21→**12 (the a11y floor)**.
+- `archive-v1` KEPT (corrected my own "dead code" framing — it's a documented governed archive
+  per CLAUDE.md; its 10 !importants already outside the live build).
+- **PUSHED both remotes** (origin Gitea + github-backup) at `40629e8`; **DEPLOYED** CloudFlare
+  + GitHub Pages mirror; **live-verified** on animalrationiscapax.org (@layer present, 12
+  !important, 34139 B). Roadmap W3R.1 marked COMPLETE (`40629e8`).
+
+### 5. Kerievsky → Chamber: HELD (background agent, completed)
+Spawned a background agent to convert the Kerievsky PDF (Desktop) to the Chamber. It
+**correctly HELD**: the PDF is an Internet Archive Scribe scan with degraded OCR — narrative
+chs 1–5 clean prose, but **code listings materially garbled** + catalog chapters suffer
+marginal-sidebar bleed. For a refactoring book the code is load-bearing. The agent used the
+existing conversion pipeline (`docs/conversion-skill-plan-2026-05-17.md` + `scripts/ocrmac_pdf.py`
++ `_curation/` — which I didn't know existed), committed NOTHING, left trail records. Prime
+directive working as intended (honest degradation over a shipped lie). Trail committed to
+chamber-library `a01bc69` (per-author-quirks Kerievsky entry + failure-notes), **unpushed**.
+Steward: **source a clean EPUB** (yes) → then the clean pipeline runs as Fowler did.
+
+## Present — mood / returns (ledger merged)
+- **Measure-don't-speculate paid repeatedly**: the 70→56 census (raw count over-counted dead +
+  comments); checking archive-v1's build status before deleting (it's governed, not dead);
+  reading the competing img specificities before judging glimpse "specificity-redundant" (it
+  wasn't).
+- **The honest-degradation gap was real and caught**: my delayer-diff "byte-identical" proof
+  gave false confidence (it proves CSS-text, not render); the steward's render-review found the
+  3 inversions. Recalibration: a structural-wrap byte-proof is necessary, NOT sufficient — render
+  is the gate when @layer changes the cascade basis.
+- **Returns**: (a) reverted the whole 5-layer attempt rather than whack-a-mole the 3 instances
+  — diagnosed the CLASS; (b) caught + restored 4 comments my sed over-strip mangled (token-strip
+  hits prose, same family as the census over-count); (c) corrected my own "archive-v1 = dead
+  code" framing on looking at the target.
+- **Challenge-the-framing twice**: caught that partial-@layer can't prove rendered-safety
+  (atomicity) before slamming 20 files; caught that the naive 5-layer was the bug, not the fix.
+- The steward's "deflation" at choosing base+overrides over the cathedral → reframed: the
+  elegance is the !important leaving (force→position), not the layer count; the cathedral is
+  deferred-not-forbidden. They "read it right."
+
+## Future
+
+### Pulling thread (singular)
+**W3R.2 — Stage 3, the spec-true tree.** Reorganize the partials largest-scale → smallest per
+the twelve scales (north-star §"The cascade order") so `main.scss` reads top-to-bottom as the
+framework's self-description — the legibility dream by *composition*. Reify a NEW @layer ONLY at
+each boundary that passes the promotion test (does it do cascade work?); classify the 27
+within-`base` collisions (2026-05-30 audit) as the reorder changes source order; utilities
+dissolution; %clearfix home; mix()/darken() deprecations. Byte-identical compile gate;
+render-verify where specificity resolution changes. **Opens by reading Kerievsky's framing
+chapters** (over/under-engineering, refactor-toward-not-impose) — readable from the PDF's CLEAN
+early chapters even before the EPUB lands.
+
+### Actionable resumption point (as of wrap — re-judge against what changed)
+ARC clean at `40629e8`, pushed + deployed + live; W3R.1 done. Stage 3 NOT started. First moves:
+(1) Read Kerievsky chs 1–5 framing from `~/Desktop/Refactoring_to_patterns_Kerievsky.pdf`
+(clean prose; the catalog chapters are the garbled part, not these) — or the EPUB if the steward
+sourced it. (2) Read `main.scss` top-to-bottom + the spec section structure (AldineXXI §I–§XIII
++ companion specs) and map each partial → its scale; note where it already sits right vs needs
+moving. (3) Pull the 27-collision list from `_audits/scss-cascade-audit-2026-05-30.md` —
+classify before reordering. Roadmap: `drafts/road-to-stage-g-2026-06-07.md` W3R.2.
+
+### Other open horizons (ranked)
+- **Kerievsky EPUB** — steward to source a clean EPUB → background conversion runs the clean
+  pipeline (Fowler precedent). Drop on Desktop. chamber-library `a01bc69` trail is unpushed.
+- **W3R.3** — the per-type×per-viewport verification gate for Stage 3 + (re)deploy. (Deploy
+  machinery proven today.)
+- **Then** Stage G gating items beyond W3R: W4.4 (multilingual v1.0 sign-off), Wave 5 (G1–G4
+  enforcement), W1.15 compass reveal (rides §VII.f build session).
+- **Baton items unchanged**: be (Seb's tenant-repo move for §2 images; a11y-gate pushed);
+  L1 (baton with Seb; connection-leak finding to relay; PENDING-27 with jurist); PENDING-32
+  stale-copy premise to relay to jurist.
+- **chamber-library** unpushed (a01bc69) — steward's call whether to push.
+
+### Pause statement
+I am about to be away from this. ARC is locked clean: the cascade declared, every forced
+override gone, the count at its irreducible a11y floor, pushed and live and verified, the
+north-star reconciled to what we actually built. The big remaining half — the spec-true tree —
+is staged but untouched. I don't know whether the steward will have sourced the Kerievsky EPUB,
+or whether Seb moved on be/L1. What I want to find still pulling is the TREE: main.scss read
+top-to-bottom until each partial sits at exactly the scale the spec gives it, the poem carried
+by the arrangement.
+
+### Literal question for next-Claude
+**Reading `main.scss` top-to-bottom against the spec's section structure — which partials
+already sit at the right scale, and where does the spec actually want the rest? (And: is the
+current @use order even the right starting frame for the tree, or does the spec corpus dictate
+a different sequence?)** Sub-checks: did the steward source the Kerievsky EPUB? Any Seb movement
+on be/L1 overnight?
+
+## Decisions deferred (and why)
+- **Stage 3 (the tree)** — its own session; deserves Kerievsky read first + a fresh frame.
+- **Kerievsky Chamber conversion** — HELD on degraded scan; awaits clean EPUB (do-once).
+- **chamber-library push** — steward's call (agent was told not to push; not explicitly asked).
+- **The 12 a11y !importants** — kept by contract (reduced-motion + visually-hidden); the floor.
+- **archive-v1 deletion** — NOT done; it's a governed archive, not dead code.
+- **Twelve @layers** — deferred-not-forbidden; base can split into more scales when the
+  rule-count earns it (the promotion test), via the same safe wrap.
+
+## Skill harvest (per §1.6) — verification-ladder additions (proposed)
+- **delayer-diff** — for any @layer / structural-wrap refactor: strip the `@layer X{`/`}`
+  wrappers from compiled output, diff vs the true baseline; byte-identical proves nothing
+  moved/dropped (only wrapped). Proven 3× today.
+- **CAUTION: byte-identical ≠ rendered-identical under @layer** — @layer overrides specificity,
+  so a structural-wrap byte-proof is necessary but NOT sufficient; render-review is the gate when
+  the cascade basis changes. (The 3 inversions the byte-diff missed.)
+- **CAUTION: a whole-file mechanical token-strip hits prose** — `sed s/!important//` also strips
+  the token from comments/code-refs; scope to declaration lines or verify comment integrity
+  after. Same family as census-through-truncation (count/measure, then look).
+→ append to `reference-verification-ladder.md` (its authorized home); surfaced in wrap §8.

claude/memory/session-ledger-2026-06-08.md

@@ -0,0 +1,32 @@
+---
+name: session-ledger-2026-06-08
+description: "Practice-of-return ledger maintained by /symmetria — returns, open horizons, recalibrations, authorization moves, sub-agent dialogues, bypasses."
+metadata: 
+  node_type: memory
+  type: feedback
+  originSessionId: 69ced5cd-4504-4a05-ba7a-357c77529189
+---
+
+# Session Ledger — 2026-06-08
+
+## Returns
+- 2026-06-08T07:40 — **measure-don't-speculate paid (Fowler ch2 date-range lesson, applied to ourselves).** The raw `grep -c '!important' = 70` was over-count: 10 dead (archive-v1, not in main.scss build) + 4 comment mentions of the word ("these three !importants…"). Clean live declaration count (`!important;` in built partials) = **56 — the audit number holds.** Two would-be mis-classifications caught by checking instead of assuming: (a) the visually-hidden mixin's 9 are LIVE not dead (it IS `@include`d at compass:97); (b) `_base.scss` has 3 real decls not 4. The literal-question census is answered by measuring, exactly as the chapter teaches.
+
+## Returns (cont.)
+- 2026-06-08T09:10 — **2a (5-layer whole-file wrap) REVERTED — design fault caught by steward render-review.** Class of failure: wrapping a heterogeneous partial in one layer promotes ITS low-specificity blanket rules (mobile `p{text-align:left}`, `*{animation}`, `.fragment{max-width:85%}`) above component rules that previously beat them by SPECIFICITY — because @layer overrides specificity. ≥3 instances spotted (mobile epigraph left-align, fragment width, epigraph/attribution spacing) but the class is unbounded by what the eye caught. My delayer-diff proved CSS-TEXT identity but NOT rendered identity — exactly the honest-degradation gap I'd flagged; the steward's eye was the real gate. Lesson: **role ≠ file** — a partial's name doesn't map to a cascade layer when its rules span roles. Reverted to clean 69d1c4a; local==live restored. Corrected design = base+overrides (everything in base = baseline cascade; only the !important override-blocks promoted) — provably rendered-identical, zero inversion surface. Awaiting steward design pick.
+
+## Open horizons
+- 2026-06-08T07:09 — ARC Wave 3R SCSS day. Discipline: READ Fowler Refactoring chs 1–3 (Chamber, craft_of_software/) BEFORE any SCSS edit. Two hats; byte-identical gate = green bar. !important re-census + per-survivor classification (ch3 Bad Smells lens) is the open literal question — raw grep now 70, not the remembered 56; classification untouched.
+- be §2 fonts/images (images gated on Seb tenant repo); L1 baton with Seb (connection-leak finding to relay; PENDING-27 with jurist); PENDING-32 stale-copy premise to relay to jurist; MemPalace 3.4.0 upgrade (saved plan, post-heavy-work slot).
+
+## Confidence to recalibrate
+- 2026-06-08T10:05 — **2b sed over-breadth: `s/[[:space:]]*!important//g` across whole files stripped the word from 4 COMMENTS too** (a code-ref in backticks, two prose mentions, and my own 2a layer-note "not by !important"). Caught by comparing raw-before (15/23) vs clean-decl (12/22) counts → diffed comment lines → restored all 4. Lesson: a mechanical token-strip across a whole file hits the token in comments/strings/code-refs; either scope to declaration lines or verify comment integrity after. (Same family as the 06-08 morning census over-count: the word lives in prose too.) Added to the watch-list for a verification-ladder note.
+- 2026-06-08T08:20 — Wave 3R Stage 2a structural proof is CONCLUSIVE *only* at the CSS-text level (delayered new == baseline, byte-identical, 34273 B). It does NOT prove rendered-identity: @layer overrides specificity, so the cascade *basis* changed. Residual risk = non-!important cross-layer specificity competition (a higher-spec rule in a lower layer that a lower-spec rule in a higher layer would now beat). The 56 !importants are RETAINED in 2a, so the previously-!important winners (compass-points, full-bleed img) are unaffected — risk is confined to normal declarations. Confirm via render-review on the per-type×per-viewport ladder before commit. Honest degradation: I cannot render here.
+
+## Authorization moves
+- 2026-06-08T09:40 — **Stage 2a RE-EXECUTED as base+overrides (steward-authorized; "which would Fowler do?" → Option 1 on Yagni/Speculative-Generality/small-steps grounds; cathedral deferred till earned).** `@layer base, overrides;` declared in _foundation with a written future-rule placement rule. 18 partials whole-wrapped → base; _compass whole → overrides; _frontispiece split in-place (epigraph/layout/nav → base, .frontispiece-compass → overrides); _variables/_mixins unwrapped. SCOPE REFINED mid-execution: overrides = compass-points ONLY (clean 34). The glimpse full-bleed 9 DEFERRED — they're `.glimpse img` nested in a mixed content-type block; clean promotion needs extraction or risks re-introducing the inversion class on figcaption/p; some may be specificity-redundant (removable with no layer). Verified: delayer-diff byte-identical to baseline (34273 B; nothing moved) + rendered-identity by construction (only already-!important-winning compass-points promoted; their non-!important rules have no base competitor; the 3 regression-causers all stay in base). 20 files, +86/−4. NOT committed/deployed. Awaiting steward re-review → commit. Then 2b = remove the 34 compass !importants.
+- 2026-06-08T08:20 — **Stage 2a (5-layer) executed in working tree** (steward-authorized over two turns): `@layer reset, base, layout, components, utilities` declared authoritatively in _foundation (first emitter); 20 emitting partials wrapped in their layers; _variables/_mixins unwrapped (emit nothing). 21 files changed, +64/−4. NOT committed, NOT deployed. _site rebuilt to wrapped state. Two hazards resolved pre-wrap: %clearfix @extend (single, cascade-neutral — Stage-3 note stands) + variables-emits-nothing. Awaiting steward render-review → then commit.
+
+## Sub-agent dialogues
+
+## Bypasses

claude/memory/skill-harvest-register.md

@@ -59,6 +59,22 @@ The single place proposed skills live so they don't evaporate between sessions.
 | `/wake-up` §2.b | patch | Until upstream #1665 closes: wake searches run **unscoped + post-filter by wing** (wing-scoped `mempalace_search` errors at HEAD). Remove when the feedback memory `feedback-mempalace-wing-filter-broken` is deleted. | 2026-06-05 diagnosis | PROPOSED |
 | `/wake-up` (new step or §2 check) | patch | **Wake canary**: seconds-cheap probe at wake — every MEMORY.md pointer + `[[link]]` resolves to an existing memory file; flag dead pointers and load-bearing orphans. Lesson 7 of substrate-failure-lessons-2026-06-07 (BM false-clean sync) applied to our own index, which the harness already part-loads with a warning. | 2026-06-07 evening (memory-verdicts night) | PROPOSED |
 
+## New proposals (2026-06-08 — harvested from Fowler, *Refactoring* 2nd ed. chs 1–3, awaiting steward)
+
+*Read in the Chamber before Wave 3R. Three elements are genuinely NEW to our toolkit (not already carried by the verification ladder / Symmetria); two more are heuristics worth a feedback memory. Proposing the new ones; flagging where each would land on authorization.*
+
+| Element | Kind | One-line | Where it lands | Status |
+|---------|------|----------|----------------|--------|
+| **Revert-and-redo-smaller** | Symmetria §3 flag OR ladder note | Ch1's strongest working reflex we *don't* have: when a verification/gate fails and the cause isn't immediately visible, **revert to the last green commit and redo the step smaller** — do NOT debug forward. We default to debug-forward; this is a different reflex, and it depends on commit-after-every-green (which we already do via the byte-identical gate). | Symmetria §3 (a flag) or verification-ladder (a named discipline) | PROPOSED |
+| **Two-hat commit separation** | verification-ladder entry | Name which hat each commit wears: a *refactor* commit's compiled output is byte-identical (or carries a pre-stated classified delta); a *feature/behaviour* commit is where rendered values change. We already own the byte-identical gate + delta-classification — this names the **commit-level rule** that pairs them ("never mix a structural move and a value change in one step"). The ARC `@layer` re-partition is the textbook case. | verification-ladder (formalizes what the byte-gate implies) | PROPOSED |
+| **Bad-smells → refactoring lens** | reference card OR fold into `/code-review` | Ch3's smell catalogue (Mutable/Global Data, Duplicated Code, Shotgun Surgery, Speculative Generality, Comments-as-deodorant…) as an explicit review lens. Today it classified the `!important` census cleanly (global-data smell → `@layer`). Likely too large for a new skill; candidate to fold into `/code-review`'s prompt or a one-page reference beside the ladder. | `/code-review` prompt or new reference | PROPOSED (lowest priority — may be redundant with existing review skill) |
+
+**Heuristics (feedback-memory candidates, no skill needed):**
+- **Rule of Three** (ch2): first time do it, second wince, third refactor. Defer abstraction to the third repetition.
+- **Preparatory refactoring** (ch2, Kent Beck): *"make the change easy (warning: this may be hard), then make the easy change."* The framing for sequencing Wave 3R *before* Stage G — Jessica Kerr's "drive north to the highway." 
+
+*Note: measure-don't-speculate (ch2 date-range story) is NOT proposed — already carried by the ladder (censused-routes, measure-toolchain-before-spec). Today's census re-proved it; no new instrument needed.*
+
 ---
 
 *Discipline reminder: "no harvest" is a valid outcome; do not manufacture proposals to fill the table (the inverse of Hermes's autonomous self-write). The yardstick is the steward's: did the steward have to re-explain something a skill could have carried?*