fix(tools): bias create_*/data toward structured form + fix_hint on JSON-decode failure

[romer8]

Summary

Companion server-side fix to the 2026-05-18 200s production incident.

The LLM emitted `data` as a JSON-string-literal containing malformed JSON (240-row token-by-token generation drifted mid-array). The tool body rejected with `invalid_args: data is not valid JSON` — but the envelope carried no recovery guidance, so the LLM regenerated the entire stringified array (~63s wasted) and frequently hit the same drift on retry.

Two changes, both within the existing `Union[List, str]` schema (no breaking change for models that always emit JSON strings):

1. Description bias. `create_plotly_chart` and `create_data_table` descriptions now STRONGLY PREFER the structured-array primary type and warn that the stringified form is error-prone for long arrays + silently doubles latency on parse failure.
2. `fix_hint` on JSONDecodeError. The rejection envelope now carries a `fix_hint` directing the LLM to retry with the structured form. Mirrors the existing input-validation middleware recovery pattern.

Why this works across all providers

• Ollama native tool-calling constrains tool_call argument STRUCTURE via the model's grammar. If the LLM picks the structured arm (which the description now strongly prefers), grammar-constrained generation prevents mid-array drift entirely.
• OpenAI strict:true (separate chatbox-core PR) enforces the same on the client side.
• Anthropic validates tool calls server-side against the schema.

Biasing the LLM toward the structured form is the actual fix; the previous JSON-mode plan's idea of passing schemas as `format` parameters doesn't work because schema-constrained generation cannot validate the content of a string value.

Tests

• 2 new regression tests in `test_tool_input_robustness.py` pinning the `fix_hint` contract on both tools.
• Suite: 786 → 788 passed.

Companion work

• chatbox-core PR (pending): OpenAI strict:true on tool definitions in the OpenAI adapter. Telemetry hook for measurement.
• Origin plan: `docs/plans/2026-05-18-003-feat-mcp-tool-call-json-mode-plan.md` in the firoh workspace (re-scoped 2026-05-18 after corrected understanding of Ollama `format` semantics).

Test plan

• `pytest test_mcp/` → 788 passed
• Live re-run of the 200s scenario (240-row time-series → create_plotly_chart) — confirm the LLM picks the structured form on first attempt OR recovers via the new `fix_hint` on second attempt