Skills: fold in four external-skills-review borrows; record the comparison memo

AGENTS.md

@@ -123,6 +123,12 @@ npm run format     # Prettier
   leads. When in doubt about existing behavior, read the spec. When you ship user-facing
   behavior, update the matching spec section in the same session; if spec and app
   disagree, the spec is stale — rewrite it deliberately, never drift silently.
+- **Reproduce before theorizing** — on a nontrivial bug, first build a command that goes
+  red on the exact symptom (failing test, script, headless-browser driver) and is fast,
+  deterministic, and runnable unattended; minimize the repro, then hypothesize against it.
+  Reading code to build a theory before that command exists is the failure mode. Write the
+  regression test before the fix; tag temporary debug logs with a unique prefix
+  (e.g. `[DEBUG-x7]`) so cleanup is one grep.
 - **Core-first** — for each feature, build the pure `src/core/` logic with tests before UI.
 - **Session wrap-up** — when the user signals the session is wrapping, run the review
   pass before any commit: `/doc-update` in-session first (flush unrecorded rationale),
@@ -172,6 +178,11 @@ carries — platform branches, state transitions, config-path writes, render ser
 not the strings it renders; if that logic is worth guarding, lift it into core/stores and
 test it there.
 
+Expected values come from an independent source of truth — a known-good literal, a worked
+example, the spec — never recomputed the way the implementation computes them. A
+tautological assertion (`expect(add(a, b)).toBe(a + b)`) passes by construction and can
+never disagree with the code.
+
 Component tests (happy-dom) share a harness shape: `createRoot` + `act` with
 `IS_REACT_ACT_ENVIRONMENT = true` set at module level, stores reset in `beforeEach`, and
 `vi.mock('../services/chart-renderer', …)` for anything that embeds a chart (vega-embed is