FM: Context recall injection — configurable memory retrieval for payload augmentation #40

Currently the OC plugin implements auto-recall client-side: it searches Ragger before each agent run and prepends results as a <relevant-memories> block. This should be a first-class feature of Ragger itself, available across all access modes (Chat CLI, HTTP API, MCP).

What this covers

• Ragger Chat (CLI/web): before assembling the payload for each turn, run a semantic search against stored memories using the user's message as the query; inject results into the context block alongside the system prompt. Already partially addressed by FM Phase 3 (FM Phase 3: Tiered payload assembly + context budget + shrinking #23) but not explicitly spec'd as a standalone, user-configurable step.
• HTTP API: expose recall as a parameter on /search (or a dedicated /recall endpoint) that returns results pre-formatted for context injection, so API clients don't need to implement it themselves.
• MCP: the search tool already exists; this issue covers making the injection step and its defaults configurable server-side (rather than hard-coded in each client plugin).

Config options (proposed, under [recall] or [chat])

recall_limit     = 3        # max memories injected per turn
recall_min_score = 0.30     # minimum similarity threshold
recall_enabled   = true     # global on/off

Notes

• The OC plugin (autoRecall) hard-codes limit=3, min_score=0.3. These should become server-side defaults that clients can override per-request.
• Injection format (e.g. <relevant-memories> XML block, prompt-injection guard language) should be defined once in Ragger, not duplicated across client plugins.
• Related: FM Phase 3 (FM Phase 3: Tiered payload assembly + context budget + shrinking #23) covers the internal payload budget; FM Phase 5 (FM Phase 5: Proxy mode — memory-augmented /v1/chat/completions #26) covers proxy-mode augmentation with augment_max_results/augment_min_score. This issue covers the general client-facing recall surface.