Correct MCP tool reference and close migration coverage gaps

[aezell]

Follow-up to a review of this package against the Agent Plugins 1.0.0 spec and the adjacent client-specific Sprites packages.

The manifests themselves were already conformant — plugin.json and mcp.json validate against the published 1.0.0 schemas, all nine skills satisfy the Agent Skills spec, and skill discovery relies only on immediate children of skills/. These changes fix inaccuracies in the reference content and restore guidance that was dropped during consolidation.

Accuracy

• tools.md risk classification. The Access column used three tiers, but the server derives MCP annotations from HTTP method (readOnlyHint for GET, destructiveHint for everything else). That means create_sprite, exec, checkpoint_create, service_create, and service_start all arrive annotated destructive, contradicting their write rows — and the public server README uses only read/write, so the table matched neither source. Renamed to Risk and added a section explaining the mismatch so an agent does not treat it as a bug or use destructiveHint alone to decide whether to ask permission.

• Missing parameters. tools.md is what the core skill points to "when parameter details matter," but it omitted several the skills already tell agents to use: wait_for_capacity (sprites-new says "wait for usable state"), max_results/continuation_token (sprites-list says "paginate"), exec_kill signal/timeout, service_logs lines/duration, start/stop duration, and the rules array shape. Also documents the exec env KEY=VALUE repeatable form, and that service_create.cmd is a single string with arguments in args — unlike exec's repeatable cmd.

• Network policy semantics. policy_network_update replaces the entire rule set rather than merging. safety.md only asked for confirmation before widening policy, which misses that a narrowing-intent update silently drops unrelated rules. Now requires a policy_network_get read-back and confirmation of the complete intended set.

Coverage

• references/deployment.md (new). The Antigravity package carried a deployment reference whose sync-exclusion list (.git/, dependency and build caches, .env and credential stores, large generated artifacts) had no equivalent here — compute.md and files.md cover how to transfer code but never what to leave behind. Restored and expanded with prefer-clone-over-copy guidance, run-time credentials instead of written secrets, and post-sync verification. Linked from SKILL.md and compute.md.

• Migration doc completeness. The artifact table reads as exhaustive but omitted one adjacent client package entirely and silently dropped the Antigravity deployment reference. Both are now mapped, including where the guard-script warnings and checkpoint-command heuristics landed.

• Compatible-client list. Corrected to the current five (ChatGPT and Codex, VS Code, Kiro, GitHub Copilot, Cursor), noting all five support Streamable HTTP so the single mcp.json entry reaches every one. The "keep the existing package" caveat now names all three clients that are not on the list. README also notes the spec is still a Working Draft.

Validation

python3 scripts/check_repository.py clean across 27 files; python3 -m unittest discover -s tests 7/7 passing. No network calls added.