diagnose: _POLICY_* interception matches bare key names at any nesting depth

[pbean]

What

diagnostics._scrub_policy (src/bmad_loop/diagnostics.py:427-434) tests each key against _POLICY_COUNT_KEYS, _POLICY_BOOL_KEYS and _POLICY_KEYSET_KEYS by bare key name, with no record of which section it sits in:

key = str(raw_key)
if key in _POLICY_COUNT_KEYS:
    out[f"{key}_count"] = len(value) if isinstance(value, (list, tuple, dict)) else 0
elif key in _POLICY_BOOL_KEYS:
    out[f"{key}_set"] = bool(value)
elif key in _POLICY_KEYSET_KEYS and isinstance(value, dict):
    ...

The sets are named for specific fields — extra_args/env/worktree_seed (count), commit_message_template (bool), settings (plugin-id keyset, hardened in #186) — but the match is depth-blind. Any future policy field that happens to reuse one of those names, anywhere in the tree, is silently rewritten by a rule that was not written for it.

Why it matters

This is an interception over-reach, not a leak — the failure direction is loss of diagnostic fidelity, not disclosure:

• a new foo.settings that is a list rather than a plugin table would be reduced to a sorted identifier-gated keyset;
• a new foo.env string would be emitted as env_count: 0 (the else 0 arm), which reads as "empty" rather than "not a collection";
• the reduced key is renamed (_count/_set), so a dump consumer sees a field that does not exist in Policy and cannot tell the substitution happened.

Nothing collides today — the name reuse is hypothetical, which is why this is a hardening item rather than a bug report.

Suggested fix

Make the interception section-aware: match on the dotted path (adapter.extra_args, plugins.settings, scm.commit_message_template, and so on) rather than the bare leaf name, threading the path through the recursion in _scrub_policy. That also makes the sets self-documenting about which section each rule was written for.

Context

Found while writing the key-passthrough invariant tests for #202, and deliberately left out of scope there (#202 changes no runtime behavior — it is a comment plus tests). Related: #186, which hardened the plugins.settings interception this rides on.