feat(mcp): detect and withhold a tool-call-free reasoning-loop reply

[alpibrupa]

Summary

• A live run today (bar mode via a "thinking" model, OpenCode's `kimi-k2.7-code`) burned its whole step budget on a ~36,000-character reply that was pure self-narration with zero real tool calls. `bridge.outcome_of_steps` returned that text as a normal, completed reply — a caller had no signal the run had degenerated.
• Scoped to lex-code rather than lex-llm: true mid-generation cancellation would need lex-llm to expose an abort, and is moot for MCP anyway — `make_handler` already calls `iter.to_list` on the whole `run_loop` before any `Step` is visible here, so the full (wasted) generation has already happened by the time detection could run.
• `is_degenerate_ramble(steps)` flags a turn with zero `StepToolExec` steps whose final text exceeds 8,000 characters. `make_handler` now checks this before handing off to `bridge.outcome_of_steps`, replacing a flagged reply with an explicit error naming the character count and pointing at `LEX_CODE_PROVIDER_` (Per-mode provider defaults / escalation for reliability-sensitive modes #116) as the fix.

Fixes #114

Test plan

• `lex check`/`lex fmt --check` — ok
• Full repo sweep — all pass
• `lex test` — 4/4 passed
• `lex doc-sync --check` — current
• Constructed-case verification covering the three cases that matter: a long zero-tool-call reply (flagged), a short zero-tool-call reply (not flagged — avoids a false positive on legitimate short chat answers), a long reply that did call a tool (not flagged — tool use is itself evidence of real work).
• Live sanity check: a normal short MCP request end to end still passes through unchanged, confirming no regression on ordinary replies.

🤖 Generated with Claude Code