feat(memory): make per-response memory recall count user-configurable (#4948)

[maxmilian]

Summary

The number of extended (non-pinned) memories retrieved into the model context per response was hardcoded to k=3 in ChatProcessor.build_context_preface, so users had no way to recall more memories (long-running projects where many small facts matter) or fewer (leaner, cheaper prompts) without editing source. This adds a "Max memories per response" control under Settings → Memory, wired through the existing per-user prefs system (/api/prefs/memory_recall_count) with a global default in settings.py. It mirrors the existing skill_max_injected setting exactly. Default stays at 3, so existing installs behave identically; the backend read clamps to [1, 50] so a malformed pref can't silently disable recall (k<=0) or balloon the prompt.

Target branch

• This PR targets dev, not main. All PRs land in dev; main is curated by the maintainer at each release. If your PR is on main by accident, click "Edit" on this PR and change the base.

Linked Issue

Fixes #4948

Type of Change

• Bug fix (non-breaking — fixes a confirmed issue)
• New feature (non-breaking — adds new behaviour)
• Breaking change (changes or removes existing behaviour)
• Refactor / cleanup (behaviour unchanged)
• Documentation only
• CI / tooling / configuration

Checklist

• I searched open issues and open PRs — this is not a duplicate.
• This PR targets dev
• My changes are limited to the scope described above — no unrelated refactors or whitespace changes mixed in.
• I actually ran the app (uvicorn app:app) and verified the change works end-to-end. Type-checks and unit tests are not enough. (Walkthrough documented in the comment below; screenshot of the rendered control included under "Screenshots / clips".)

How to Test

Automated (added with this PR):

python -m pytest tests/test_memory_recall_count_4948.py -q

• test_mem_recall_count_drives_retrieval_k — a non-default count is forwarded to the retrieval step as k.
• test_mem_recall_count_defaults_to_current_behavior — omitting it preserves the previous hardcoded k=3.

Manual (reviewer):

1. Run the app (uvicorn app:app or docker compose up) and open Memory → Settings.
2. Confirm a new "Memory recall — Max memories per response" number input (default 3).
3. Change it (e.g. to 10); it persists to /api/prefs/memory_recall_count (verify with GET /api/prefs/memory_recall_count).
4. Chat with several saved memories; more/fewer extended memories are pulled into the preface as the value changes.

Visual / UI changes — REQUIRED if you touched anything that renders

• Screenshot or short clip of the change in the running app — attached below (Settings -> Memory, the new "Memory recall" card with "Max memories per response" set to 10).
• Style match: reuses the existing admin-card layout, admin-toggle-sub labels, and the same <input type="number"> styling as the adjacent Inject Skills control — no new color values, font sizes, spacing units, or classes introduced.
• No new component patterns. Extends the existing memory settings panel and reuses the shared syncPrefNumber helper (added an optional toast-label arg, default unchanged) rather than writing a parallel one.
• I am not an LLM agent submitting a bulk PR. Single focused feature against a pre-existing, labeled issue (Make the number of recalled memories per response user-configurable #4948); happy to adjust to maintainer preference (presets vs. raw number, "Unlimited" semantics) before merge.

Screenshots / clips

[maxmilian]

One design question before I flip this out of draft, since the issue (#4948) left the control shape open:

The issue proposed two shapes — presets (Low / Default / High / Unlimited) vs. a plain "max memories per response" number input. This PR currently implements the number input: minimal, backward-compatible (defaults to the current 3, clamped to [1, 50]), and it sits right next to the existing Inject Skills → Max skills per request control in the same panel, so the two read consistently.

I leaned number-input for that consistency, but I'm glad to switch to the preset buttons if you'd prefer. If you do want presets, two sub-questions:

• Should Unlimited mean "all memories above the similarity threshold" (the issue's recommendation) rather than literally every memory, to avoid filling context with irrelevant entries?
• Any preferred default preset / values for Low and High?

Happy to rework the UI either way before review — just let me know which shape you want. (Backend is the same either way: the count flows through to the recall k.)

[maxmilian]

Following up on the design question above — going with the number input as implemented, for the consistency reasons noted (it sits beside the existing "Max skills per request" control and reads the same way), and it's minimal + backward-compatible (defaults to the current 3, clamped [1,50]). Flipping out of draft now; happy to switch to the preset buttons during review if a maintainer prefers — the backend is identical either way.