Agent Behaviour: stop character leaking into Memory; migrate preference:* and responseStyle

[serge-ivo]

Depends on #223.

Why

The Coder wrote preference:response_style into Memory, and it is still there. Once #223 lands there
are two places holding the same idea, and the older one is the one the agent already knows how to
reach.

Scope

1. Prompt — the memory section states plainly that Memory is for subject-matter knowledge and
   that character/tone/verbosity live in behaviour, with set_behaviour named as the way to change
   them. Without this the model keeps using the tool it already knows.
2. Migrate existing preference:* memory entries into config.behaviour where they map cleanly;
   leave the rest and let the user see them.
3. responseStyle free-text → structured. Guardrails.responseStyle is a free string
   ("professional" | "casual" | "concise" | …) read in exactly two places. "technical" maps to a
   technicality band; the rest map to tone. Keep reading the old field for instances that have it.

Separate bug found in the same log

write_memory fired four times with the same key in a single turn. The cross-round dedup of
identical (name, args) calls should have collapsed that. Harmless here (last write wins) but it
means the dedup is not catching what it claims to — worth its own look while in this code.

[serge-ivo]

The main scope looks shipped (daed127 — agent-behaviour.ts, behaviour-store.ts, set_behaviour/get_behaviour in BASE, Behaviour tab wired in surfaces.tsx), so items 1–3 appear done.

Not closing, because of the sub-bug recorded at the bottom of this issue: write_memory fired four times with the same key in one turn, which the cross-round (name, args) dedup should have collapsed. Nothing in the behaviour commits touches that dedup, so it is still open and now buried inside a closed-looking ticket. Worth splitting out into its own issue before this one is closed, or it will be lost.

[serge-ivo]

Closing as implemented on main. Evidence: commit daed127 moves character out of Memory: agent-think warns that behaviour belongs in set_behaviour, agent-behaviour.ts detects stray preference:* behaviour keys, and tests cover the migration/read path. Follow-up dedup work also landed in de2505f.

[serge-ivo]

The sub-bug recorded at the bottom of this issue is fixed in de2505f, so it is no longer a reason to hold this open.

write_memory firing four times against one key was not the dedup failing to run — it was the dedup keying on JSON.stringify(args), which is key-order dependent. The same logical call with its fields emitted in a different order hashes differently and walks straight past the check, so nothing was deduplicating identical intent, only identical byte-order.

Fixed by keying on a sorted-key stringify. The pump's delivery idempotency key had already solved this; it is now lib/stable-json.ts and shared, rather than a second copy that could drift from the first.

Items 1–3 of this issue appear shipped in daed127. Close when you have confirmed that yourself — I did not want to close another session's work on inspection alone.

[serge-ivo]

Implementation links: no associated PR was found; this landed direct on main in daed127, with related follow-up in de2505f.