Summary

Adds is_suggestion + rationale (and confidence) params to the granular update_issue_assignees MCP tool, so agents can propose assignees as pending suggestions rather than applying them directly.

Why

Closes https://github.com/github/plan-track-agentic-toolkit/issues/333

The other granular tools (update_issue_type, update_issue_labels, set_issue_fields) already support intent params behind the issues_granular feature flag. update_issue_assignees was the last granular tool still on the plain string-array form. This brings it to parity, following the same implementation pattern as the labels tool in #2557.

What changed

• Reworked GranularUpdateIssueAssignees from the simple issueUpdateTool wrapper to the object-form pattern: assignees items are now oneOf: [string, object], where the object is { login (required), rationale (≤280 chars), confidence (low/medium/high), is_suggestion }.
• Handler parses string-or-object entries, validates rationale length and confidence enum, and sends either the plain IssueRequest{Assignees:} (no intent) or an object-form body ({ login, rationale, confidence, suggest }) via a raw PATCH — mirroring update_issue_labels.
• Added assigneeWithIntent and assigneesUpdateRequest structs.
• Regenerated the tool snapshot and docs/feature-flags.md.
• Added integration tests: object-form (plain / suggest / rationale+confidence / mixed) and validation errors (rationale too long, missing login, invalid confidence).

MCP impact

• Tool schema or behavior changed — update_issue_assignees now accepts an object form per assignee with rationale / confidence / is_suggestion. The plain string-array form is unchanged and still serializes identically, so it's backwards compatible.

Prompts tested (tool changes only)

Tested locally against dotcom via a local build (bin/github-mcp-server-local, --features issues_granular):

• Suggest an assignee with rationale + high confidence → issue returned, assignee not applied (pending suggestion created), confirmed via the API.
• Invalid confidence value → tool error confidence must be one of: low, medium, high.
• Malformed entry (123) → tool error each assignee must be a string or an object….
• Empty assignees array → clean no-op.
• Non-assignable login with is_suggestion: true → dotcom 422 surfaced as a clean tool error.

Security / limits

• No security or limits impact — same auth/scopes (repo) and endpoint as before; only the request-body shape gained optional intent fields. Rationale is capped at 280 chars.

Tool renaming

• I am not renaming tools as part of this PR

Lint & tests

• Linted locally with ./script/lint
• Tested locally with ./script/test

Docs

• Updated (README / docs / examples) — regenerated via script/generate-docs.