v2.4.0-unsloth-studio-lift: port 3 Unsloth Studio AGPL-3.0 modules

Batch 1 — tool-call-parser.ts: replaces xml-parser.ts with a port of
Unsloth's tool_call_parser.py. Adds balanced-brace JSON scanner,
single-param fast path, hasToolSignal/stripToolMarkup/parseToolCallsFromText
exports, and stream-finalization stripping at all three final-write sites
(error-handler, finalizeCompletion, executeToolPhase). Anthropic <invoke>
shape preserved. 75+12 tests.

Batch 2 — web/html-to-md.ts: parse5 tree-walking HTML-to-Markdown converter
ported from Unsloth's _html_to_md.py. Replaces web_fetch's regex stripHtml
with structured markdown output (headings, links, lists, tables, code blocks,
blockquotes, entity decoding). 29 tests.

Batch 3 — llama-args-validator.ts: port of llama_server_args.py deny-list
validator. Wired into AGENTS.md frontmatter parser — llama_extra_args field
validated at load time, rejects managed flags (model identity, networking,
auth/TLS, server UI). No runtime consumer yet (llama-swap boundary). 76 tests.

All three files carry SPDX-License-Identifier: AGPL-3.0-only headers.
LICENSE flipped to AGPL-3.0-only in prior commit (a938cf1).

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

apps/server/src/services/inference/tool-phase.ts

@@ -14,6 +14,7 @@ import { formatUnknownToolError } from './tool-suggestions.js';
 // Resolves the grant root before pausing the loop so the user is never
 // prompted about paths we couldn't grant anyway (e.g. /etc/passwd).
 import { resolveGrantRoot } from '../grant_resolver.js';
+import { stripToolMarkup } from './tool-call-parser.js';
 import type {
   InferenceContext,
   StreamResult,
@@ -100,7 +101,8 @@ export async function executeToolPhase(
   projectRoot: string
 ): Promise<ToolPhaseResult> {
   const { sessionId, chatId, assistantMessageId } = args;
-  const { content, toolCalls, promptTokens, completionTokens } = result;
+  const content = stripToolMarkup(result.content, { final: true });
+  const { toolCalls, promptTokens, completionTokens } = result;
 
   // v1.11.3: ctx_max comes from llama-swap /upstream/<model>/props, not the
   // streaming completion (which doesn't emit n_ctx). getModelContext caches