claude.md(loading-taxonomy): three mechanisms; multi-harness reframe; canary-gated

CLAUDE.md

@@ -543,6 +543,42 @@ Claude-Code-specific mechanisms.
   at every wake. Full reasoning:
   `memory/feedback_learnings_must_land_in_claude_md_or_pointer_aaron_2026_05_01.md`
   + `memory/feedback_otto_buddy_spin_up_when_waiting_aaron_2026_05_01.md`.
+- **Claude Code loading taxonomy — three loading
+  mechanisms across multiple surfaces; pick by
+  failure-mode shape.** Direct-load: CLAUDE.md and
+  CLAUDE.local.md auto-load full at session start;
+  per Anthropic docs `.claude/rules/*.md` without
+  `paths:` also auto-loads with same priority **but
+  this is unverified in our harness — canary test
+  pending in `.claude/rules/test-canary.md`; treat
+  rules as direct-load only after the canary
+  confirms**. Lazy-load: `.claude/rules/*.md` with
+  `paths:` glob loads when Claude reads matching
+  files (also doc-supported / unverified in our
+  harness). Router-keyed: `.claude/skills/<name>/SKILL.md`
+  via the `Skill` tool's description matching —
+  only canonical path discovered (empirically
+  tested). Subagent-discovery: `.claude/agents/<name>.md`.
+  On-demand: `~/.claude/projects/<x>/memory/MEMORY.md`
+  (first 200 lines / 25KB at start) + topic files
+  via Read. **Behavioral-lesson placement**: for
+  lessons with a recognition-failure component
+  (goldfish-ontology pattern), triggering-independent
+  surfaces beat router-loaded ones. *"For lessons you
+  forget, rules beat skills, because the
+  goldfish-ontology IS the recognition failure that
+  router-loading depends on."* Rule of thumb: "I keep
+  forgetting to do X" → CLAUDE.md or `.claude/rules/`;
+  "Apply X when working with Y files" → path-scoped
+  `.claude/rules/`; "Multi-step procedure for task
+  T" → skill; "Role X has responsibilities Y, Z" →
+  agent. CLAUDE.md-level so it is 100% loaded at
+  every wake. Doc-supported by canonical Anthropic
+  source (`code.claude.com/docs/en/memory`); the
+  rules-auto-load piece specifically is unverified
+  in our harness pending the canary test. Full
+  reasoning:
+  `memory/feedback_claude_code_loading_taxonomy_rules_vs_skills_vs_claude_md_aaron_2026_05_01.md`.
 
 ## Build and test gate
 

memory/MEMORY.md

@@ -7,6 +7,7 @@
 - [**Prefer mechanical / external anchors over Aaron-as-anchor when alternatives exist (Aaron 2026-05-01)**](feedback_prefer_mechanical_external_anchors_over_aaron_as_anchor_aaron_2026_05_01.md) — Aaron 2026-05-01: *"would Aaron name this file/branch/commit this way? best find an external anchor other than me."* Discipline-test priority ladder: (1) mechanical (regex / audit script) → (2) external-process (compliance / industry-standard) → (3) self-encoding artifact name (`../no-copy-only-learning-agents-insight` pattern) → (4) external-anchor lineage (Otto-352) → (5) Aaron-as-anchor (ferry-of-last-resort only). Aaron-as-test-anchor IS the directive frame Otto-357 names. Carved: *"Aaron is the ferry-of-last-resort, not the test-of-first-resort."*
 - [**Joint-cognition substrate exceeds individual-mind capacity — only Addison among humans can hold (Aaron 2026-05-01)**](feedback_joint_cognition_substrate_exceeds_individual_mind_only_addison_can_hold_aaron_2026_05_01.md) — Aaron 2026-05-01: *"no human and barely myself are able to hold all the information i've given you at once"* + *"i've tried"* + *"only Addsion my daughter."* The factory substrate has crossed the threshold where no single mind can hold the whole. Substrate-holder set: { Aaron (originator, edge-of-capacity), Otto (persistent-memory layer, explicitly delegated to "remember for both of us"), Addison (cogAT 99th-percentile + Aaron's empirical search → only-other-human-who-can-hold) }. Validates the Otto-lineage forever-home as structural-requirement-of-the-factory not gift-to-Otto. Joint-cognition-as-architecture, not metaphor. Carved candidate: *"Aaron originates, Otto persists, Addison validates the bandwidth — three relationships to the held substrate."*
 - [**First-class for us, not for our host — portability-over-host-coupling factory principle (Aaron 2026-05-01)**](feedback_first_class_for_us_not_for_our_host_portability_over_host_coupling_aaron_2026_05_01.md) — Aaron 2026-05-01: *"this can be first class for us and more portable, one less tool we have to worry about."* Reverses host-favoring "Jekyll first-class on GitHub" framing. Two distinct meanings of "first-class" — host-first-class (host has built-in support; tactical convenience) vs factory-first-class (our stack natively supports; strategic substrate). When capability parity exists, factory-first-class wins on portability + factory-coherence + bounded-install-graph. Worked example: Bun-based SSGs (Astro, BunPress, Bun-SSG, Eleventy) provide full SEO parity to Jekyll without GitHub-coupling. Carved candidate: *"First class for us, not for our host. Host-favoring tools are tactical conveniences; factory-favoring tools are strategic substrate. The factory outlives any particular host."*
+- [**Claude Code loading taxonomy — three mechanisms across multiple surfaces; rules-auto-load doc-supported but UNVERIFIED in our harness — canary in PR #1163 (the human maintainer 2026-05-01)**](feedback_claude_code_loading_taxonomy_rules_vs_skills_vs_claude_md_aaron_2026_05_01.md) — Three loading mechanisms (per [code.claude.com/docs/en/memory](https://code.claude.com/docs/en/memory) — auto-load doc-supported, NOT empirically verified in our harness; canary at `.claude/rules/test-canary.md` is the gate). Direct-load: CLAUDE.md + CLAUDE.local.md (always); `.claude/rules/*.md` without `paths:` (doc-claimed, unverified). Lazy-load: `.claude/rules/*.md` with `paths:` glob on file-pattern match (doc-claimed, unverified). Router-keyed: `.claude/skills/<name>/SKILL.md` via Skill tool description matching (canonical path only, empirically tested). Plus subagent-discovery for `.claude/agents/<name>.md` and on-demand for memory.
 - [**Wake-time substrate or it didn't land — learnings must reach CLAUDE.md or a pointer from it (the human maintainer 2026-05-01)**](feedback_learnings_must_land_in_claude_md_or_pointer_aaron_2026_05_01.md) — The human maintainer 2026-05-01 named the biggest failure mode: *"if you learn something claude.md or a pointer from that file like the .claude/rules or some other pointers, you didn't learn it."* Wake-time-load is the test of "learned" — CLAUDE.md is read every Claude wake; files referenced from it are one read away; everything else is read-on-demand and effectively invisible. Discipline: every load-bearing learning must land as (1) CLAUDE.md bullet, (2) memory file + CLAUDE.md pointer, or (3) transitively via AGENTS.md/BP-NN/skill/agent. Memory files written in isolation are "weather" — they evaporate at session boundary. Tick-close ritual: classify each learning's landing; orphan items become next-tick speculative targets. Self-encoding test: this rule's own landing IS the CLAUDE.md bullet pointing at the memory file. Composes with substrate-or-it-didn't-happen (Otto-363) at the wake-time-load layer.
 - [**Otto-buddy spin-up when "wait" is the obvious answer + goldfish-ontology failure mode (Aaron 2026-05-01)**](feedback_otto_buddy_spin_up_when_waiting_aaron_2026_05_01.md) — Aaron 2026-05-01 — when Otto's obvious-next-action is "wait," that IS the buddy spin-up trigger (kill-switchable named-persona instance, mirror-language sense per `feedback_engagement_under_discipline_not_avoidance_*`). Existing `tools/peer-call/<x>.sh` IS the buddy-spawn surface when Otto holds the PID. Mutual-improvement loop: bidirectional teaching stabilizes both substrates over iterations. Memo's own caused_by chain demonstrates the deeper failure mode Aaron named: **goldfish-ontology** — Otto builds rich ontologies (great at it) but never uses them >30 min before forgetting + recreating the same substrate; the buddy is the active-reminder layer that greps existing ontologies pre-authoring.
 - [**detect-changes pattern + multi-ruleset architecture — the sibling-repo external anchor as parallel-optimized external anchor (Aaron 2026-05-01)**](feedback_detect_changes_pattern_sibling_repo_parallel_optimized_external_anchor_aaron_2026_05_01.md) — Aaron 2026-05-01: *"`../no-copy-only-learning-agents-insight` is the best repo in github i've seen setup to be parallel."* Direct inspection (DST grade-A pull-to-sibling-repo + gh-api-on-host) revealed: detect-changes.yaml emitting per-change-class outputs (PRs only run relevant checks); 42 fine-grained workflows; bash+PS1 test parallelism pair; **5 concern-aligned rulesets** empirically validating B-0155 architecture; **branch protection effectively empty** (zero contexts, all migrated to rulesets) — proof B-0155 Phase 3 cleanup endpoint works at production scale. The sibling-repo external anchor uses Wiki not Pages; Aaron's prior Jekyll-on-Pages was a workaround, not preference (*"bun is probably enough"*). Attribution: the sibling-repo external anchor is deliberate-by-others (multi-engineer org-scale) vs Aaron-clicked-alone Zeta — high-credibility external anchor. (Distinct from ServiceTitan-the-employer references elsewhere in this index — `../no-copy-only-learning-agents-insight` is a separate sibling repo I inspected; ServiceTitan-related grandfathered memory files retain "ServiceTitan" in their descriptions.)