fix(mcp): consolidate and recall memory on every request

[alpibrupa]

Summary

• `make_handler` called `ag.run_loop` directly on a raw `AgentLoop`, never constructing a `session.lex` `Session` — so `consolidate.run()` never ran for MCP requests. Any `remember` tool call made over MCP wrote a candidate to `.lex/memory-candidates.jsonl` that stayed an unconsolidated candidate forever, regardless of provider.
• Since brains are built once in `main()`, before any candidate exists, a per-request refresh is the only place MCP can pick this up. Each `code` tool call now:
  1. Consolidates whatever candidates are pending under a fresh synthetic id (`crypto.random_str_hex(16)` — MCP has no client-tracked session to attribute them to).
  2. Re-recalls current memory and prepends it to the selected brain's system prompt via `sess.with_memory` (same mechanism `session.lex` uses for ACP/TUI/graph.lex), so a fact remembered on one call is visible — attested in the trail — by the next.
• Recall goes through `pmem.open()` + `recall_for_prompt()` rather than the higher-level `recall_context()`: the latter's `fs.exists` guard needs `fs_walk`, which `Skill.handle`'s fixed effect row (`lex-agent/src/server.lex`) does not carry. `open()` already creates the db lazily on a missing file, so the guard isn't needed here.

Fixes #109

Test plan

• `lex check src/server/mcp_main.lex` — ok
• `lex fmt --check` — formatted
• Full repo sweep (`lex check` over every tracked `.lex` file) — all pass
• `lex test` — 4/4 passed
• `lex doc-sync --check` — current
• Live pipeline test with a standalone driver replicating this exact code path (`candidates.propose()` → `consolidate.run()` → `pmem.open()`/`recall_for_prompt()`): candidate correctly promoted into `project_memory.db` with a full attestation trail in `memory_trail.db`, and recall rendered `"\n\nMemory:\n- [convention/commit-style] ..."` — which `prefix_goal`'s `str.trim` correctly treats as non-blank and prepends to the agent's goal