feat: robust strict mode implementation

[VforVitorio]

What problem does this solve?

strict mode is intended to be a pure chat experience with no tool access. The current implementation might be leaky or not cleanly strip tools from the LM Studio SDK payload, confusing the model or leading to hallucinated tool calls.

Proposed solution

1. When in strict mode, intercept the tool list passed to model.act() and send an empty list (or cleanly disable tools at the SDK level).
2. Update the system prompt dynamically if needed so the model knows it cannot interact with the local environment.
3. Ensure UI elements reflect this (e.g., hiding tool-related hints or panels).

Alternatives considered

N/A - Strict mode needs to strictly mean 'no tools', requiring SDK-level blocking.

[VforVitorio]

Fixed in commit f85107e on branch feat/robust-auto-strict-modes.

Root cause: _run_turn in src/lmcode/agent/core.py was building the tool list unconditionally (tools = [self._wrap_tool(t) for t in self._tools]) and passing it to model.act() regardless of self._mode. The mode label said "no tools — pure chat only" but the SDK still received every tool schema, so a model in strict mode could emit tool_calls and the runtime would execute them silently.

Fix: Strict is now enforced at the SDK boundary — when self._mode == "strict", tools=[] is passed to model.act(). The model literally cannot see the tool schemas. No prompt engineering needed, no extra guard rails; the SDK protocol itself refuses tool_calls for tools it doesn't advertise.

Regression tests added in tests/test_agent/test_core.py:

• test_run_turn_strict_mode_sends_empty_tools
• test_run_turn_non_strict_modes_pass_tools (ensures ask/auto are not collaterally affected)

Also fixed the README description which claimed strict was "read-only" (it was never read-only).

Will ship as part of the auto+strict PR once #97 is also done.

[VforVitorio]

Correction to the previous comment — the fix described above (tools=[] passed to model.act()) turned out not to work: the LM Studio SDK rejects an empty tools list with LMStudioValueError, so strict mode crashed as soon as the user sent a message.

The fix that actually landed, in #103, has two layers:

1. SDK layer — strict mode routes through model.respond() (the pure-chat SDK primitive) instead of model.act(). respond() has no tool concept at all, so the model never sees a tool schema. This is the real SDK-level enforcement.

2. Prompt layer — model.respond() stops the model from emitting real tool calls, but cannot stop it from fabricating tool output in plain text. Qwen 7B, for example, happily replied "Here is the file content:" followed by invented code because the base system prompt said "Always call read_file before describing file contents". Strict mode now passes a separate lms.Chat seeded with a dedicated _STRICT_SYSTEM_PROMPT that explicitly forbids pretending to read files, fabricating output, or describing tool calls — while still listing what the model can do (review pasted code, answer general questions, rubber-duck debugging) so strict stays useful.

self._chat keeps the base prompt so Tab-switching back to ask/auto retains full history without a rebuild. Both layers have regression tests in tests/test_agent/test_core.py.