fix(copilot-on-rails): require quoted mermaid labels in debug plan diagrams

[Nathan (nturinski)]

Base branch: feat/CoR — not main. The Copilot on Rails code does not exist on main.

Problem

The azure-debug-plan agent generated an ## Architecture Diagram in .azure/vscode-debug-plan.md containing this edge:

API -->|@azure/storage-blob| AZ

Mermaid v11 reserves @ for edge-ID syntax (e1@-->) and node-metadata syntax (id@{ shape: ... }), so a @ at the start of an unquoted label token is lexed as LINK_ID and the whole diagram fails to parse:

Parse error on line 15:
...|pg| PG    API -->|@azure/storage-blob
---------------------^
Expecting 'AMP', 'COLON', 'PIPE', 'TESTSTR', 'DOWN', 'DEFAULT', 'NUM', 'COMMA',
'NODE_STRING', 'BRKT', 'MINUS', 'MULT', 'UNICODE_TEXT', got 'LINK_ID'

LocalPlanView's MermaidBlock catches the throw from mermaid.render() and falls back to a raw code block labeled mermaid (error) — so the user sees diagram source instead of the diagram.

Verified behavior (headless, jsdom, real mermaid builds):

Diagram	Result
A-->|@azure/storage-blob| B	❌ fails
A[@azure/storage-blob]-->B	❌ fails — node labels affected too
A-->|uses @azure/storage-blob| B	✅ passes — @ mid-label is fine
A-->|"@azure/storage-blob"| B	✅ passes — quoting fixes it

Only a leading @ breaks. Scoped npm package names (@azure/storage-blob, @azure/identity, @prisma/client) are exactly what the agent puts in Azure-dependency edge labels, so this is very reachable in practice.

Not a regression

Reproduced identically on 11.14.0 (the version originally added in #1436), 11.15.0 (the dependabot bump in #1439), and 11.17.0 (what resolves today for "mermaid": "^11.15.0"). The LINK_ID grammar predates the CoR mermaid dependency entirely.

package.json and package-lock.json are deliberately untouched — bumping or pinning mermaid would not fix this.

Other characters in the failing diagram (<br/>, the · middle dot, the [( )] cylinder shape, :5173 port suffixes, blank lines between statements) all parse fine and were ruled out.

Fix

Harden the agent-facing authoring guidance so generated diagrams always quote labels.

• resources/agents/azure-debug-plan/references/plan-template.md — new ## ⚠️ Mermaid Label Quoting — ALWAYS Quote Every Label section alongside the existing ## Markdown Table Integrity rule (outside the ```markdown template fence, so it reads as an agent rule rather than template body). It carries a ✅/❌ example pair and a concise explanation of the LINK_ID cause. Inline %% hints were also added inside the template's Architecture Diagram block so the rule is visible at the exact point of authoring, not just in a section the model might skim past.
• .github/instructions/copilot-on-rails-docs.instructions.md — the same rule attached to the existing "keep the Mermaid pipeline diagram accurate" bullet, since that file also instructs an agent to maintain a mermaid diagram (the docs/copilot-create-project.md pipeline flowchart).

The rule is unconditional — quote every node label and every edge label, always — rather than an @-specific special case. Quoting is always valid in mermaid and additionally protects (, ), :, /, #, and ·. Asking the model to detect which labels are "risky" is far less robust than a rule it applies every time.

Intentionally not changed: resources/agents/azure-project-plan.agent.md and resources/agents/azure-project-plan/plan.md mention mermaid only to prohibit it in the parsed project plan, so a quoting rule there would be wrong.

Lockfile note

evals/agent-assets.lock.json is refreshed via node evals/check-agent-drift.ts --update, because plan-template.md is a hash-tracked agent asset and editing it fails the drift gate. Only that file's hash, the rollup agentAssetsHash, and updatedAt changed — no scope or tracked-file-set change.

Validation

The full credential-free contracts CI job was run locally against fc079fd; all 12 gates pass:

drift · typecheck · imports:self-test · certify (152/152 grader certification cases) · stacks:check · gates · phases:check · seed:contract · lint · clean-machine:check

lint reports 3 pre-existing scoring-defaults-applied warnings that are unrelated to this change.

Parse behavior was confirmed directly against real mermaid builds in a throwaway sandbox outside the repo: the unquoted form fails, the quoted replacements parse. The four existing fixture plans under evals/grader-certification/** and test/testProjects/copilotOnRails/** were each checked and already parse cleanly today (none has a leading @), so they were left alone rather than churned.

The behavioral MSBench evals are workflow_dispatch-only (they need Entra auth, not GITHUB_TOKEN) and are being dispatched separately against this PR.

[Nathan (nturinski)]

Copilot resolve the merge conflicts in this pull request

[Copilot]

Copilot resolve the merge conflicts in this pull request

Merged feat/CoR in (70a267b) and resolved both conflicts: LocalPlanView.tsx now keeps the onRenderStatus callback from feat/CoR while using the label-quoting renderMermaid fallback, and agent-assets.lock.json was regenerated via node evals/check-agent-drift.ts --update. tsc --noEmit, the webview build, and the drift check all pass clean.