Intent drift on shell tool: model plans exec_command but hook enforces Bash (register-and-retry loop)

[Harihara04sudhan]

Problem

On Codex, a first tool use triggers a frustrating register-and-retry loop:

1. Agent reads a file -> PreToolUse blocks: `ArmorCodex intent mismatch: parameters not allowed for Bash`
2. Agent registers a plan with `action: "exec_command"` (Codex's API tool name)
3. PreToolUse blocks: `ArmorIQ intent drift: tool not in plan (Bash)` -- because Codex reports the shell tool as `Bash` at the hook layer, and `normalizeToolName` only lowercases, so `exec_command != bash`
4. Agent re-registers with `action: "Bash"` -> finally allowed

This wastes turns and makes the plugin feel like it's fighting the user.

Root cause

• plugins/armorcodex/scripts/lib/common.mjs normalizeToolName only trims+lowercases; no alias between Codex's shell tool names (exec_command/shell/local_shell/unified_exec) and Bash.
• The register_intent_plan directive (engine.mjs) says "action = tool name" without telling the model the canonical names, so it guesses exec_command.

Fix

• Alias the Codex shell family to bash in normalizeToolName so a plan declaring any of them matches the actual Bash tool call (and policy rules for bash cover the shell family).
• Make the register_intent_plan directive name the canonical tools (Bash for shell, apply_patch for edits, exact MCP tool name), and tell the agent to register all tools up front with metadata.inputs: {}.
• Tests in tests/tool-alias.test.mjs.

Refs

Part of the ArmorCodex parity work (#41). Fix on branch feat/armor-policy-parity.