fix(cursor): keep the line range when bridging the native read tool

[QuangBlue]

Problem

Cursor models read files with their native read_args exec, which the executor bridges to the client's declared read tool. The decoder kept only ReadArgs.path and dropped offset (field 4, int32) and limit (field 5, uint32), so every partial read reached the client as a whole-file read.

Claude Code answers a repeated whole-file Read of an unchanged file with "Wasted call — file unchanged since your last Read". The model never got the lines it asked for and retried the same read indefinitely. On our deployment, 1877 of 2306 Cursor-routed Read calls over three days got that answer, and one session read the same file 342 times. It is worst after Claude Code's own auto-compaction, when the file content is gone from context but Claude Code still considers it read, so a ranged read is the model's only way back to it.

Wire capture of the failing turn: exec_read path=…/service.py offset=1195 limit=16 was forwarded to the client as Read {"file_path": "…/service.py"}.

Fix

• cursorAgentProtobuf: decode offset (sign-extended int32) and limit (uint32) on exec_read.
• builtinToolBridge.readBridge: forward them when the client's read tool declares offset/limit as integer/number properties whose bounds (minimum, maximum, exclusiveMinimum, exclusiveMaximum) accept the values. A read tool that cannot carry the range is not a match, so a ranged read keeps the typed rejection instead of silently becoming a whole-file read.

Tests

• Decoder: offset/limit kept; negative offset decoded as int32.
• Bridge: Claude Code's Read schema (verbatim, integer + bounds), a number-typed read tool, a path-only tool (not bridged), out-of-bounds values (not bridged), unranged read unchanged.

Reproduced and verified against a real Cursor model through a local router with Claude Code (claude -p): before, 2/3 runs looped to the turn limit with 9–11 "Wasted call" results; after, 3/3 answered correctly with none.

node --import tsx/esm --test tests/unit/*cursor*.test.ts passes (the sigkillFollowupMs timing test is flaky under load and passes on its own). check-open-sse-typecheck is clean.

⚠️ base-red inherited: #15032 (check-file-size flags tests/unit/response-sanitizer.test.ts on the base tip; this PR does not touch it).