feat!: carry Standard Schema issues on InvalidEntity

[btravers]

Reworked per review. The previous approach flattened issues into
"field: message" strings; this keeps them structured.

Why

fromSchema already returns SchemaIssues (readonly StandardSchemaV1.Issue[])
and toInvalidEntity was discarding it. Flattening was lossy — a message can
itself contain ": ", so the path could not be recovered — and it forced a
caller to re-parse a string the library had just serialised from data it held.
@unthrown/standard-schema is already a peer dependency, so adopting its
Issue type costs nothing.

It was also self-inconsistent: instance already propagated structured issues
into zod's channel with real paths.

Changes

• errors.ts — issues: SchemaIssues instead of readonly string[].
• entity.ts — toInvalidEntity passes issues through untouched. construct
  wraps invariant messages as { message } with no path.
• instance.ts — forwards each issue's own path into ctx.addIssue, so a
  nested field failure now reports ["owner", "slug"] rather than just
  ["owner"]. Two tests pin this.
• issues.ts (new) — keysOf (Standard Schema permits both bare PropertyKey
  and { key } segments; zod emits the bare form) and renderIssue, used
  only for the add-defect Error message, which has nowhere to put
  structure.

Behaviour

ApiKey.decode({ ...raw, secret: "short" });
// issues: [{ path: ["secret"], message: "Too small: …" }]

Trial.decode(brokenRow);
// issues: [{ message: "trialEndsAt must be after createdAt" }]   ← no path

z.object({ owner: Organization.instance }).safeParse({ owner: { slug: "" } });
// issues: [{ path: ["owner", "slug"], message: "Too small: …" }]

The pathless-invariant case is new information, not just a different encoding:
a whole-entity rule failure is now distinguishable from a field issue that
happened to have an empty path.

Gate

format --check, lint, typecheck (both passes), test (86, 9 files),
knip, build — all green. Rebased onto current main.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.