Harden companion bridge against top conversational pain points

[RobVanProd]

What changed

• Added a source-linked corpus of 100 companion-robot and conversational-agent complaints, then ranked the normalized clusters through a four-judge council.
• Ran three bridge-only hardening iterations covering relationship safety, repetition, STT uncertainty, shared-room privacy, persona-scoped episodic memory, corruption recovery, initiative control, model/TTS failure recovery, and research-grounding failure behavior.
• Added an executable top-20 qualification gate that maps every ranked complaint cluster to a named control and concrete regression tests.
• Added deterministic and configured-Gemma red-team evidence.

Why

The bridge was conversationally capable but several failures could still break trust or continuity: model and TTS faults ended sessions, low-confidence speech could be treated as understood, episodic callbacks were not persona-scoped, a search outage could give the model room to improvise a fresh fact, and proactive-speech preference persistence was not host-owned.

This PR makes those boundaries explicit and testable before physical-robot qualification.

User impact

• Previous played context survives a failed model turn and capture reopens.
• TTS failure preserves the conversation lease and reopens after the acoustic tail.
• Low-confidence STT requests a repeat without writing memory or invoking tools.
• Public research outages return an immediate honest local response without a wasted model generation.
• Persona episodes and callbacks cannot bleed across persona packs.
• Spoken initiative opt-out is immediate, persisted, reversible, and host-owned.
• Dependency pressure, harmful sycophancy, and repetitive full replies are repaired before speech.

Validation

• python -m unittest discover -s bridge -p "test_*.py": 543 passed.
• python bridge/companion_harness_qualification.py --run: 32 passed, 20/20 controls, no skips.
• Configured gemma4:e2b-it-qat Character Lock red team: 29/29 passed, zero validation failures and zero runner errors.
• git diff --check: clean.

Scope and residual limits

• Host bridge only. No firmware, actuator, or motion behavior changed.
• STT confidence gating applies when the configured adapter supplies confidence; the current whisper.cpp path may omit it.
• Shared-room suppression uses aggregate observed person_count; this does not claim speaker identity.
• Physical-robot conversational qualification remains the final release check.
• The unrelated local face GIF modification is not included.