since-cutoff 0.4.1: removed parameters point to extra_body; imported-scope notes lead with what breaks

[MohammadHijjawi97]

since-cutoff 0.4.1 is on PyPI (uvx since-cutoff@latest scan). Two changes to the notes, both found while preparing the head-to-head benchmark.

A removed parameter no longer reads as "drop it"

For anthropic 1.8.0 the 0.4.0 note said: "Messages.create() no longer accepts temperature, top_k or top_p; do not pass them." In the benchmark pilot, a task asked for temperature=0. Every agent run with that note obeyed it and left the temperature out, so the hidden test failed. The parameters left the signature, but the pinned create() still sends extra fields through extra_body. A static diff can see that; it cannot see whether the API still accepts the field. The note now says only what it knows:

Messages.create() no longer accepts temperature, top_k or top_p as keyword arguments. If the API still needs them, pass them through its extra_body or extra_query argument. [diff]

This applies to any SDK method that takes extra_body / extra_query: anthropic, openai and the other Stainless-generated clients. Other removed parameters keep "do not pass them".

sync --scope imported leads with what breaks new code

With --scope imported (notes for the packages your code imports, not only the APIs it already uses), the notes now come in this order:

1. import paths that no longer work;
2. changes with a known replacement;
3. removed classes and functions;
4. removed parameters and members;
5. deprecations.

• Whole packages that moved are now one note. mcp 2's mcp.server.fastmcp → mcp.server.mcpserver used to be one note per module.
• Renamed classes are now found by comparing members. FastMCP → MCPServer is found because 29 of its 30 public members are kept; on anthropic, openai, huggingface-hub and pandas this finds no renames at all.
• --per-package N sets how many notes per package.

On mcp 1.28.1 → 2.2.0 the five notes went from parameter removals on ClientSession and friends to: the package move, the two renames (McpError → MCPError, FastMCP → MCPServer), streamablehttp_client → streamable_http_client, and list_prompts(cursor=) → params.

1013 tests, 94% branch coverage. Full changelog: https://github.com/MohammadHijjawi97/since-cutoff/blob/main/CHANGELOG.md

Benchmark status. The pilot (2 tasks × 4 arms × 2 runs, 16 sessions) is done and excluded from the analysis. It confirmed the harness works: Context7 connected in 8 of 8 runs, and a session took a median of 54 s at a median cost of $0.26. In the pilot, the docs arm called Context7 in only 2 of 4 runs. The pre-registered rule for that case is a fifth arm that is instructed to use Context7, so the main run is now 24 tasks × 5 arms × 3 runs. Because the note wording changed after I saw a pilot task, the results will be published twice, with and without the two pilot tasks, and a claim counts only if it holds in both.