Keep spoken memory silent in TTS replies

[tylerreckart]

Arthur (and any mode: "spoken" / channel: "voice" agent with /mem) was reading memory machinery aloud: entry titles, ids, bodies, search hit lists, “I found in memory…”. Memory should inform the reply silently.

Ban list (spoken prose)

Silent /mem writs are unchanged (recall-first, write-as-you-go). The listener must never hear:

• Entry ids (#42, id 12)
• Titles as labels (Coffee preference:)
• Type names as spoken categories (user / feedback / …)
• Search rankings, hit lists, or “top match”
• “Searching memory”, “I found in memory”, “according to my notes/memory”
• Verbatim entry bodies, [TOOL RESULTS] / /mem search dumps
• Graph dumps, neighbour lists, /mem expand / density output
• Listing multiple hits for the listener to pick from

Teach: answer as if you simply know. If they asked you to remember, one short confirmation only (I'll keep that.).

How

• Hypothesis A: prompt_spoken_memory() Speech rules now hard-ban the list above and last-win over research COMMAND RULES (print /mem entry, show the graph).
• Hypothesis B: Spoken COMMAND RULES skip Honeycomb-style graph examples and defer to MEMORY HABIT so those dumps cannot dominate TTS. Text/TUI mem UX is unchanged.
• Hypothesis C (post-filter / render_policy) not used — prompt fix only; HTTP/SSE and Intercom contracts unchanged.

Verify

1. ctest --test-dir build -R unit_constitution --output-on-failure
2. Ask Arthur something that triggers /mem search (a stored preference, a person, “what did we decide”). The spoken reply should use the fact, not read the entry, cite an id, or say it searched memory.
3. Ask Arthur to remember something: one short confirmation, plus a silent /mem add entry writ — do not hear the title/body.

See Voice → Personal-assistant memory.