fix(schema): align file_write mode default with Python implementation

[shaun0927]

Problem

assets/tools_schema.json (used for English-locale agents since the i18n
auto-detect in 7cadbd7) declares file_write.mode with a default of \"append\":

"mode": {"type": "string", "enum": ["overwrite", "append", "prepend"],
         "description": "Write mode", "default": "append"}

But assets/tools_schema_cn.json and the Python handler both default to
\"overwrite\":

// tools_schema_cn.json line 35
"default": "overwrite"

# ga.py:369
mode = args.get("mode", "overwrite")

When the LLM trusts the English schema's documented default and omits
mode, the Python handler silently overwrites the file. Same call,
opposite behavior depending on which locale loaded the schema.

Fix

Change the EN schema default to \"overwrite\" so all three sources of
truth — EN schema, CN schema, Python handler — agree:

-      "mode": {"type": "string", "enum": ["overwrite", "append", "prepend"], "description": "Write mode", "default": "append"}}}
+      "mode": {"type": "string", "enum": ["overwrite", "append", "prepend"], "description": "Write mode", "default": "overwrite"}}}

The reverse direction (changing CN + ga.py to \"append\") was rejected
because:

• The Python handler has shipped overwrite since the introduction of
  the parameter, so overwrite is the de-facto behavior callers depend on.
• file_write in CONTRIBUTING-style usage (huge edits with the
  <file_content> block before the call) targets a fresh write, not an
  append; flipping the default to append would silently change semantics
  for every existing skill that omits mode.

Verification

$ python -c "import json; json.load(open('assets/tools_schema.json'))"  # parses
$ diff <(grep -A0 '"default"' assets/tools_schema.json) <(grep -A0 '"default"' assets/tools_schema_cn.json)
# defaults match for file_write.mode

Net line count: 0.