chore(infra): align pyright (LSP) with ty (validation) on MCP server scope

[dougborg]

Problem

pyright (used by the IDE/LSP via pyrightconfig.json) and ty (the project's actual type checker, via poe typecheck) disagree about whether statuspro_mcp_server/ is in scope:

• pyrightconfig.json includes statuspro_mcp_server/ — the LSP surfaces type warnings on every edit there.
• pyproject.toml poe typecheck runs ty check --exclude 'statuspro_mcp_server/' … — so those same warnings never block validation, never gate commits, never fail CI.

Net effect during today's MCP refactor work (PRs #51, #52): pyright noise lit up MCP files mid-edit, but couldn't actually be acted on because ty (which is what agent-check / check runs) wasn't checking the same files. Either the warnings are real and we should be enforcing them, or they're noise and the LSP shouldn't be surfacing them.

Options

1. Bring MCP under ty — drop the --exclude 'statuspro_mcp_server/' and fix whatever surfaces. Aligns the two checkers; eliminates the silent-divergence trap. Likely some real fixes needed since the directory has been ungated for a while.
2. Quiet pyright on MCP — add statuspro_mcp_server to pyrightconfig.json's exclude. Cheaper but leaves a directory with no static type checking at all.
3. Switch the LSP to ty — once ty matures its LSP story (currently pre-alpha per the comments in pyproject.toml). One checker, one source of truth. Probably the long-term answer, but not today.

Recommend Option 1: align by tightening, not by loosening.

Acceptance

• poe typecheck runs against statuspro_mcp_server/ with no errors
• pyrightconfig.json and the ty task list match (or document why they intentionally diverge)
• No new files added to either exclude list as a workaround

Context

Surfaced during PR #51/#52 refactor work. CLAUDE.md already calls out "ZERO TOLERANCE for ignoring errors" — an excluded directory is a quieter form of the same thing.

[dougborg]

Concrete data point from today's session that strengthens Option 1 (bring MCP under ty):

While tightening MCP tool parameter annotations in statuspro_mcp_server/src/statuspro_mcp/tools/orders.py (adding Annotated[…, Field(description=…)] wrappers), pyright surfaced two classes of real type errors that ty was silently missing:

1. Pydantic Field(None, …) not recognized as defaulted — three call sites of BatchOrderResult(...) were flagged "Argument missing for parameter order_id" because pyright doesn't infer default=None from the positional form. Real fix was switching to Field(default=None, …) in schemas.py.
2. str | None → str | Unset mismatch — UpdateOrderStatusRequest(comment=comment, …) and BulkStatusUpdateRequest(comment=comment, …) were passing str | None from the tool param into attrs fields typed str | Unset. Real fix was wrapping with to_unset(comment) per CLAUDE.md's documented None → UNSET convention.

Both classes were latent — they'd been incorrect for as long as the code's been there — and ty check --exclude statuspro_mcp_server/ never reported them. Pyright caught them only because the LSP scope diverges from the validation scope (this issue's premise). Without #56's resolution, future MCP edits will keep accumulating these silently.

Also re-confirmed the noise side: pyright emits "is not accessed" hints on every @mcp.tool-decorated inner function (the local def name is registered by side-effect, never read). With Option 1 those go away because ty doesn't surface that hint. With Option 2 (quiet pyright) you trade the noise for the silent-divergence trap demonstrated above — strictly worse.

Recommend keeping the recommendation as Option 1.