Update CLAUDE.md with test harness and all 44 MCP tools

- Added sx_harness_eval to the MCP tools table
- Added spec/harness.sx to the specification files list
- Added full test harness design section (sessions, interceptors, IO log,
  assertions, extensibility, platform-specific extensions, CID-based
  component/test association)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

CLAUDE.md

@@ -98,6 +98,12 @@ Read the error carefully. Fragment errors give the parse failure in the new sour
 | `sx_diff_branch` | Structural diff of all `.sx` changes on branch vs base ref |
 | `sx_blame` | Git blame for `.sx` file, optionally focused on a tree path |
 
+**Test harness:**
+
+| Tool | Purpose |
+|------|---------|
+| `sx_harness_eval` | Evaluate SX in a sandboxed harness with mock IO. Returns result + IO trace. Params: `expr`, optional `mock` (SX dict of overrides), optional `file` (load definitions) |
+
 **Analysis:**
 
 | Tool | Purpose |
@@ -168,6 +174,8 @@ The SX language is defined by a self-hosting specification in `shared/sx/ref/`.
 - **`shared/sx/ref/primitives.sx`** — All ~80 built-in pure functions: arithmetic, comparison, predicates, string ops, collection ops, dict ops, format helpers, CSSX style primitives.
 - **`shared/sx/ref/render.sx`** — Three rendering modes: `render-to-html` (server HTML), `render-to-sx`/`aser` (SX wire format for client), `render-to-dom` (browser). HTML tag registry, void elements, boolean attrs.
 - **`shared/sx/ref/bootstrap_js.py`** — Transpiler: reads the `.sx` spec files and emits `sx-ref.js`.
+- **`spec/harness.sx`** — Test harness: mock IO platform for testing components. Sessions, IO interception, log queries, assertions (`assert-io-called`, `assert-io-count`, `assert-io-args`, `assert-no-io`, `assert-state`). Extensible — new platforms add entries to the platform dict. Loaded automatically by test runners.
+- **`spec/tests/test-harness.sx`** — Tests for the harness itself (15 tests).
 
 ### Type system
 
@@ -212,6 +220,26 @@ lambda, component, macro, thunk (TCO deferred eval)
 
 The `aser` (async-serialize) mode evaluates control flow and function calls but serializes HTML tags and component calls as SX source — the client renders them. This is the wire format for HTMX-like responses.
 
+### Test harness (from harness.sx)
+
+The harness provides sandboxed testing of IO behavior. It's a spec-level facility — works on every host.
+
+**Core concepts:**
+- **Session** — `(make-harness &key platform)` creates a session with mock IO operations
+- **Interceptor** — `(make-interceptor session op-name mock-fn)` wraps a mock to record calls
+- **IO log** — append-only trace of every IO call. Query with `io-calls`, `io-call-count`, `io-call-args`
+- **Assertions** — `assert-io-called`, `assert-no-io`, `assert-io-count`, `assert-io-args`, `assert-state`
+
+**Default platform** provides 30+ mock IO operations (fetch, query, action, cookies, DOM, storage, etc.) that return sensible empty values. Override per-test with `:platform` on `make-harness`.
+
+**Extensibility:** New platforms add entries to the platform dict. The harness intercepts any registered operation — no harness code changes needed for new IO types.
+
+**Platform-specific test extensions** live in the platform spec, not the core harness:
+- `web/harness-web.sx` — DOM assertions, `simulate-click`, CSS class checks
+- `web/harness-reactive.sx` — signal assertions: `assert-signal-value`, `assert-signal-subscribers`
+
+**Components ship with tests** via `deftest` forms. Tests reference components by name or CID (`:for` param). Tests are independent content-addressed objects — anyone can publish tests for any component.
+
 ### Platform interface
 
 Each target (JS, Python) must provide: type inspection (`type-of`), constructors (`make-lambda`, `make-component`, `make-macro`, `make-thunk`), accessors, environment operations (`env-has?`, `env-get`, `env-set!`, `env-extend`, `env-merge`), and DOM/HTML rendering primitives.