feat(prompts): parameterized workflow prompts + server instructions + project prompt files

[dehuaichendragonplus]

Why

The built-in MCP prompts were canned one-liners: all declared zero arguments, and prompts/get ignored its input and returned a single role:user message that just restated the prompt's own description — no real tool names, no parameters, no guidance. They were effectively undiscoverable workflow stubs.

Two more gaps:

• Cross-client conventions lived only in CLAUDE.md/AGENTS.md (Claude Code-specific). A fresh MCP client had no protocol-native way to learn how to drive this Unity server correctly (structured envelopes, Play Mode / domain-reload sequencing, etc.).
• A project could not register its own prompts without forking the package, so project-specific workflows (activity/build/port recipes) had nowhere to live at the protocol layer.

What changed

1. Built-in prompts → parameterized workflow templates. edit_prefab_safely, verify_compilation, enter_play_and_recover, wire_serialized_references, create_playable_prototype each now declare real arguments, and prompts/get expands into an ordered sequence of actual tool calls with the caller's arguments interpolated, plus a Guidance section on that workflow's pitfalls. Declared-required arguments are enforced (a missing one returns a clear message instead of a body full of unfilled placeholders). A single declarative PromptSpec table is the source of truth for names, arguments, and the embedded resource.

2. Live resource embedding. Prompts that act on current state embed the relevant read-only resource inline in prompts/get — verify_compilation → unity://errors/compilation, wire_serialized_references → unity://selection/current, create_playable_prototype → unity://scene/active — so the model sees that state without a separate fetch. Best-effort: silently skipped if the resource provider is unavailable or the read fails.

3. Server-level instructions. The initialize result now returns an instructions string with the cross-cutting conventions a client needs (structured {success,…} envelopes and branching on code; never hand-editing serialized assets; the request_recompile / Play Mode / get_reload_recovery_status sequence; instanceId reuse via find_method=by_id; group_duplicates logs; small-payload screenshots). This reaches every MCP client, not just Claude Code.

4. Project-authored prompts. A project can drop mcp-prompts/*.md files in its root to register its own workflow prompts — no fork, no C#. Format: a --- front-matter fence (name, description, arguments: a(required), b) + a body template referencing {arg} placeholders. They appear in prompts/list alongside the built-ins and interpolate arguments in prompts/get. Reserved built-in names cannot be shadowed; malformed files are skipped with a logged warning (not fatal). Parsing is dependency-free (no YAML library).

Benefits

• Prompts become genuinely useful, self-contained workflows a model can select and follow — with the real tool names and the pitfalls spelled out — instead of stubs.
• Server instructions fix "conventions only Claude Code knows": every client gets them at handshake.
• "Generic workflows in the package, project workflows in the project" — projects extend prompts without forking, which keeps this package generic.

Compatibility

• Additive. MCPPromptProvider's new constructor parameter is optional (resourceProvider = null), so existing construction still compiles; initialize gains one field; no tool signatures change.
• Prompt names changed (fix_compile_errors→verify_compilation, runtime_validation→enter_play_and_recover, auto_wire_references→wire_serialized_references, plus new edit_prefab_safely). Prompts are user-selected, not a hard API; README updated. Flagging it since it is a visible rename.
• No new package dependency (front-matter parsing is hand-rolled; reflection-free).
• instructions is part of the initialize result in protocol 2024-11-05 (the version this server declares), so no protocol-version bump is needed.

Dependency / sequencing

The edit_prefab_safely prompt and one line of instructions reference set_prefab_property, the stage-free prefab-edit tool proposed in #45 (fix(prefab): edit prefab assets without prefab-stage save corruption). This PR is text-only with respect to that tool — it compiles, and all tests pass, on the current main without #45 — but the guidance names a tool that only exists once #45 lands. Recommended merge order: land #45 first, then this PR; or merge this one knowing the edit_prefab_safely single-field path references a not-yet-present tool until #45 merges. Nothing else in this PR depends on #45.

Testing & verification

• Verified on a live Unity 6000.3.13f1 editor via the MCP tools: prompts/list returns the reworked built-ins with their arguments; prompts/get interpolates arguments into the real tool sequences; a project mcp-prompts/*.md file is picked up and interpolated; reserved built-in names are not shadowable.
• 22 new EditMode tests across MCPServerInstructionsTests, MCPPromptProviderTests, ProjectPromptLoaderTests: instructions content, prompt argument declarations + interpolation + required-arg enforcement, live-resource embedding, front-matter parsing (valid / missing name / missing fence / arg spec), folder scan + malformed-skip, provider merge, and the built-in-name-cannot-be-shadowed guard. All pass; no regressions in the rest of the EditMode suite.

Changelog

Added under ## Unreleased.