fix(jac-gpt): don't crash startup when MCP tool categorization LLM call fails

[udithishanka]

Problem

The jaseci test-pypi-build CI job (Run tests for jaseci) has been red on main since jaseci-labs/jaseci#7042, failing at "Wait for jac-gpt server to be ready" (example run). The jac-gpt server never binds its port and the wait step times out.

That job scaffolds jac-gpt from the published jacpack (generated from this jac-gpt-fullstack/) and runs jac start main.jac.

Root cause

services/jacServer.jac made a blocking by llm() call (categorized_mcp_tools) at module / glob-init time to categorize the MCP tools, gated on the MCP client having discovered tools:

glob mcp_tools_cache: dict[str, list[str]] =
    reorder_categorized_mcp_tools(categorized_mcp_tools(available_mcp_tools_descriptions))
    if available_mcp_tools_descriptions else { ... };

This call was dormant until #7042: jac-gpt has no jac-mcp dependency, so before the fold jac mcp was an unknown command, McpClient got 0 tools, available_mcp_tools_descriptions was empty, and the else branch ran (no LLM call). #7042 made jac mcp a built-in command, so the client now connects, discovers 19 tools, and the by llm() call fires at startup.

Two further problems made that call fragile:

• it ran before load_dotenv() and before glob llm = Model(...), so it used the default model with credentials not yet loaded;
• any LLM error (no key / unreachable, as in CI) aborted module load, taking down the whole server.

The "missing credentials" error is not jaseci dropping the key: the jac-gpt CI step has no real LLM key, and byllm reads the key from the environment at call time. This is a jac-gpt startup-ordering / resilience bug that #7042 merely exposed.

Fix

• Load .env up front (own with entry) so credentials are available before any LLM call.
• Move categorization into a build_mcp_tools_cache() helper wrapped in try/except that falls back to the empty cache. The MCP chat abilities already read it via .get(..., []), so an empty cache is safe and the feature degrades gracefully.
• Initialize the cache after glob llm is configured, so the intended model is used when a key is present.

Net effect: the server always boots; tool categorization is best-effort and only runs when both MCP tools and a working LLM are available.

Verification

Reproduced and verified against a jac binary built at a red commit (f6aed53e, which includes #7042):

• Before: by llm() at glob-init with no key -> AuthenticationError -> Error loading <module> (server never starts), matching the CI trace.
• After: same conditions (no key / invalid key) -> failure is caught, cache falls back to empty, and the module finishes loading ("server booted").
• jac check passes on the changed constructs; jac format parses the file cleanly.

Once merged to main, the Generate jac-gpt Jacpack workflow republishes the jacpack to jaseci-labs/jacpacks, which unblocks the jaseci CI job.