MCP tool description truncation at 2048 chars is silent, unconfigurable, and can leave a misleading prefix (follow-up to #41593, #81268)

## Summary

MCP tool `description` fields are silently truncated to 2048 chars before being assembled into the model's context. There is no log line, no setting to override it, and the `/mcp` UI shows the *untruncated* text — so an author verifying their tool description in `/mcp` sees content the model never receives. Worse, for tools whose description embeds a usage contract (e.g. a TypeScript interface for a generic "execute" tool), the surviving 2048-char prefix can look complete and closed while critical fields are documented only past the cut — the model then confidently uses a wrong, narrower API with no signal anything is missing.

This is a follow-up to #41593 and #81268, adding a controlled same-day comparison and a new "misleading prefix" failure mode neither prior issue covered.

## Verification

Binary: `/Users/victorb/.local/share/claude/versions/2.1.234` (Claude Code v2.1.234, macOS darwin 25.5.0).

```
LC_ALL=C grep -aoE '.{160}Tpe\b.{80}' /Users/victorb/.local/share/claude/versions/2.1.234 | head
```

The minified constant `Tpe=2048` gates it. The MCP tool wrapper contains (verbatim from the binary):

```js
async description(){return W},async prompt(){return W.length>Tpe?ei(W,Tpe)+"… [truncated]":W}
```

`description()` — surfaced in the `/mcp` UI — returns the full string `W`. `prompt()` — what actually gets assembled into the model's context — truncates it. So the UI and the model see different text, with no indication of that divergence anywhere in the UI.

Server `instructions` truncation does log a debug line ("Server instructions truncated from X to 2048 chars"). Tool `description` truncation logs nothing anywhere.

## Impact

Controlled comparison on one MCP server (an OpenAPI code-execution MCP server whose `execute` tool embeds its full usage contract in the tool description), same server, same day:

- **claude.ai** receives the full description: **16,275 chars**.
- **Claude Code** (including Cowork)

receives **2,061 chars** (2,048 + the 13-char `"… [truncated]"` marker) — **12.7%** of the document.

Verbatim captures of the same server's tool schemas as received by Claude Code (truncated) and by claude.ai (complete) are attached to this issue.

The surviving prefix contains a TypeScript `RequestOptions` interface listing `method, path, query, body, contentType, rawBody` that *looks* complete and closed — but `headers`, `bodyBase64`, `multipart`, and `returnAs` are documented only past char 2048 (see attachments). The model has no way to know the interface it's looking at is incomplete, so it confidently calls the tool with a narrower, wrong API. None of this is logged, and `/mcp` shows the author the full, untruncated text, so the author has no way to notice the model is working from a different (and misleading) document.

## Prior issues

- #41593 — "MCP tool descriptions truncated at 2KB breaks code execution tools". Same diagnosis, on an earlier build (then `WoH=2048`, v2.1.84). Closed by the stale-bot with no decision.
- #81268 (open) — "MCP truncation at 2048 chars is invisible". Documents the no-log and `/mcp`-mismatch aspects; reports 502 truncation events on one machine, worst case 10,780 → 2,048 chars.

This issue adds: a controlled claude.ai-vs-Claude-Code comparison on the same server/day, and documentation of the misleading-closed-interface failure mode (the truncated prefix isn't just short, it's shaped like a complete spec).

## Requested fixes

1. A server-settable per-tool override, e.g. `_meta["anthropic/maxDescriptionChars"]`, mirroring the existing `_meta["anthropic/maxResultSizeChars"]` pattern already read by the binary (`O._meta?.["anthropic/maxResultSizeChars"]`) for result-size overrides. Keeps the tight default while letting code-execution tools opt in to a longer description.
2. Log every tool-description truncation the same way server-`instructions` truncation is already logged.
3. Make `/mcp` display the truncated text the model actually receives (or at minimum mark the cut point), so authors aren't debugging against a document the model never saw.
4. Alternatively/additionally: a `maxToolDescriptionLength` setting in `settings.json`.

## Environment

- Claude Code v2.1.234
- macOS darwin 25.5.0

## Attachments

[toolsearch_results_gmail_ba_codemode_sessionB.html](https://github.com/user-attachments/files/31182966/toolsearch_results_gmail_ba_codemode_sessionB.html)
[toolsearch_results_gmail_ba_codemode.html](https://github.com/user-attachments/files/31182967/toolsearch_results_gmail_ba_codemode.html)

Start with the reported Claude Code v2.1.234 binary and the MCP tool wrapper's description() and prompt() paths; reproduce the 2048-character difference and inspect how /mcp renders descriptions. Compare this with server-instructions logging and the _meta maxResultSizeChars handling. Done should include an agreed treatment for truncation, visibility, and configuration, with verification that the model and /mcp show consistent information.

Do mô hình lập chỉ mục viết ra từ nội dung của issue.