test: cover CLI core loop, MCP tools, and richRender branches

[benvinegar]

What

Adds 53 tests across the three highest-risk coverage gaps identified by node --experimental-test-coverage. No source changes — tests only.

Why

A coverage audit (274 tests, all green) surfaced three gaps ranked by product invariant, not raw line count:

1. bin/sideshow.js (CLI) — a first-class integration tier ("agents with only a shell can use this") sat at 63.7% lines / 44% branches. The core publish → comment → revise loop — the product's central flow — was entirely untested.
2. server/mcpHttp.ts (MCP) — another first-class tier. upload_asset was untested, plus get_design_guide and the JSON-RPC error paths.
3. server/richRender.ts — builds the sandboxed HTML served from /s/:id under a sandbox CSP header (a load-bearing isolation invariant). 61.5% branch coverage with zero unit tests (only e2e).

The feedback cursor / exactly-once delivery (the #1 guarded invariant) was already well covered — no action needed there.

Changes

test/cli.test.ts (+38 tests, +527 lines)

The CLI core loop, pinned to a real in-process server via SIDESHOW_URL/SIDESHOW_SESSION:

• publish — file + stdin (-), combined surfaces (--md/--code/--terminal/--mermaid), --code language inference from filename, --json invalid-JSON error, --diff --layout split
• diff / markdown / code (with --line-start/--filename/--language) / terminal (--cols/--term-title) / json (valid + invalid) / mermaid
• update — revise to v2 on the same card; missing-id usage error
• wait — returns a pending comment; timedOut on no comments; bad --after fails fast
• comment — --author override, --surface back-compat alias, missing --post usage error
• list / list --all / sessions
• image / upload / asset-url (content hash, no server hit)
• error paths: unreachable server (one-line error, no stack trace), server error surfaced

test/api.test.ts (+11 tests, +136 lines)

MCP /mcp endpoint, appended to the existing MCP section:

• upload_asset — base64 round-trip (id + url + kind + retrievable bytes), missing-data error, explicit-session attach
• get_design_guide returns the guide text
• publish_post empty-surfaces error
• update_snippet legacy html argument revises to v2
• JSON-RPC: malformed body → −32700 (400), batch → −32600 (400), notification → 202, ping → {}

test/richRender.test.ts (new, +14 tests, 117 lines)

First unit tests for the server-side rich-surface renderers:

• markdown: link target=_blank + rel=noopener noreferrer, code highlighting
• code: empty-language plain fallback, unknown-language try/catch fallback, lineStart counter-reset, shiki highlight path, filename header bar, copy-button JS-string escaping (\u003c)
• diff: explicit part.files (no patch), multi-file patch (one <diffs-container> per file), processFile fallback for headerless patches, empty-patch throws
• terminal: ANSI → colored spans, cols width, default "terminal" title

Coverage

Area	Lines	Branches	Functions
Overall	94.2% → 96.8%	85.6% → 87.3%	90.9% → 93.3%
bin/sideshow.js	63.7% → 78.9%	44.0% → 45.7%	54.6% → 66.3%
server/mcpHttp.ts	85.1% → 100%	67.1% → 80.7%	81.8% → 81.8%
server/richRender.ts	92.7% → 100%	61.5% → 87.5%	78.3% → 95.8%

Remaining CLI gap is platform-specific agentPid process-tree walking (Windows/ps) and the trace-sync/demo/watch loops — lower-risk than the core loop.

Validation

• npm test — 327 pass, 0 fail (was 274)
• npm run typecheck — clean (all three tsc programs)
• npm run lint — 0 warnings, 0 errors
• npm run format:check — clean

Maintenance-only (test-only) → empty changeset.