fix: drop hardcoded planning-agent ban from GPT-5.6 delegation guard

[AndrewTilson]

What

Removes a single hardcoded line from _build_gpt_5_6_invoke_agent_guard_text() in code_puppy/agents/_builder.py:

"Never invoke \`planning-agent\`. "

Why

This string is appended to the system prompt of any agent whose model name contains gpt-5.6 (substring match via _is_gpt_5_6_family). The result is that GPT-5.6 agents refuse to delegate to planning-agent and explain the refusal as a runtime prohibition.

That explanation is misleading, because no such runtime block exists:

• _invoke_agent_impl in code_puppy/tools/subagent_invocation.py refuses only on depth grounds (_subagent_recursion_blocked / _gpt_5_6_recursion_blocked). There is no agent-name allow/deny list anywhere in the invocation path.
• planning-agent remains registered and visible in list_agents output, so users see an agent that is listed, appears invocable, and is nonetheless refused.

The guard text is also appended last — after the system prompt, puppy rules, and the extended-thinking note — which maximises its salience for exactly the models that treat terse imperatives most literally.

Worth noting the inconsistency with its immediate neighbour: the sibling depth cap in the same function is user-tunable via the subagent_recursion_limit_gpt_5_6 config key and carries a comment explaining why the limit is read at prompt-assembly time. This agent-name ban had no config escape hatch and no stated rationale.

Scope

One line deleted. The depth cap and the shell-safety overlay are untouched.

Testing

• python -m py_compile code_puppy/agents/_builder.py — passes
• pytest tests/test_subagent_invocation_usage.py — 51 passed
• No test in tests/ asserts on the removed string

Note for reviewers

If this ban was originally added to work around a specific GPT-5.6 failure mode (for example a planning-agent delegation loop), then removing it outright is the wrong fix and it should instead become a documented, config-gated option. I could not find a comment or commit message stating such a rationale, so this PR takes the straightforward reading. Happy to convert it to a config key if you know the backstory.

Ref: https://jira.walmart.com/browse/PUP-665

[WSxDemise]

[Wes's CodePuppy Agent Review]

Removes the GPT-5.6 prompt-level ban on planning-agent while retaining the configured depth cap. Scope is limited to the requested guard text; no user-facing strings or comments changed. PUP-665 AC4 remains incomplete because no guard-text test was added or updated. Qodo: not configured for OSS/public GitHub.

What you're being asked to change on this PR

• [P0] [AC NOT MET] Add coverage for the changed GPT-5.6 guard — code_puppy/agents/_builder.py:548

[WSxDemise]

[P0] [AC NOT MET] Add coverage for the changed GPT-5.6 guard

• Why: The removed prohibition can return without a regression signal, leaving PUP-665's required guard-text coverage unmet.
• Do: Add a focused test asserting the GPT-5.6 guard preserves the depth cap and omits the planning-agent prohibition.

Evidence

PUP-665 AC4 requires tests covering the GPT-5.6 guard text to be updated; this PR changes only _builder.py, and no test references _build_gpt_5_6_invoke_agent_guard_text or the removed literal.