feat(memory): cap auto-recall entries at 200 chars per agent

[arcuru-bot]

What

Auto-recall surfaces memory entries into the context tail with each entry's value rendered unbounded — a single long memory could crowd out the rest of the recalled set. The 200-char-per-entry default described in docs/src/design/autonomous_memory.md was never implemented.

This adds the cap:

• New knob: max_entry_chars on the per-agent auto-recall config — default 200, settable 1–2000 via /memory config set max_entry_chars <n>, shown in /memory config show.
• Enforcement: the auto-recall context-tail path (own memory + each bank) threads the cap into search_memory/format_entry, which truncate an entry's value to the cap at whole-char boundaries with an … elision marker. Multibyte values are never split mid-character.
• Scope: the interactive recall tool keeps full-fidelity output — it is not the context-tail path. The max_entries count, bank selection, and defaults are unchanged; configs stored before this change deserialize with the 200-char default.

Tests

• Value exceeding the cap is truncated to the cap with the elision marker
• Short and uncapped values pass through untouched
• Char-boundary safety on a multibyte value
• search_memory(…, Some(20)) honors a non-default cap end-to-end through the store
• Config default/round-trip: stored configs without the key deserialize to 200

Docs

sessions.md lists only skills as a PromptAugmentation contributor — memory recall is a ContextTail contribution.

Breaking change

Follow-up rename (breaking, accepted at this stage of development): the auto-recall config keys max_entries and max_entry_chars are renamed to auto_recall_max_entries (default 3, range 1–20) and auto_recall_max_chars (default 200, range 1–2000). /memory config set with an old key now returns an unknown-key warning listing the valid keys, and stored configs with old keys silently fall back to defaults (3/200).