feat(cli): add memory introspection commands

[Teerapat-Vatpitak]

Summary

Adds vibe-trading memory list/show/search/forget so users can inspect
and manage persistent memory without opening ~/.vibe-trading/memory/
by hand. Surfaces real bug signal: two recent memory fixes (#87
snake_case, #95 CJK slugs) were hard to triage without observability
into what is actually saved and what the recall scorer returns.

Diff vs main: +516 / -5 across cli.py, src/memory/persistent.py,
and two test files. 81 tests pass. vibe-trading --version measured
at 1.0s (no regression on cold startup).

CLI surface

vibe-trading memory list   [--type user|feedback|project|reference]
vibe-trading memory show   <name>
vibe-trading memory search <query> [--limit N]
vibe-trading memory forget <name> [-y|--yes]

<name> resolves by exact title first, then by on-disk filename stem
({type}_{slug} or the bare slug), so users can paste either form
from the index.

Exit codes mirror existing commands:

scenario	code
success	EXIT_SUCCESS (0)
user declines confirm	EXIT_SUCCESS (0)
not found / EOF stdin	EXIT_USAGE_ERROR (2)
remove failed	EXIT_RUN_FAILED (1)

Commits

The branch is split into four commits so each step is reviewable on
its own:

Commit	Why
892cd16 feat(cli): add memory introspection commands	the actual feature
f8de9d0 perf(cli): avoid loading agent runtime on cli startup	hoisting PersistentMemory made --version 7.6s because src.agent.frontmatter transitively loads AgentLoop / SkillsLoader. Reverted to lazy imports inside each cmd_memory_* and a local _MEMORY_TYPES tuple in cli.py so non-memory commands stay fast
fe4fca6 style(cli): revert unintended formatter spillover on unrelated functions	a local pre-tool-edit formatter reformatted ~30 unrelated functions in cli.py. This commit restores them to upstream form so the diff only shows memory-feature lines
672e2ba fix(memory): coerce frontmatter values to str in _scan_entries	bug fix discovered during heavy E2E — see below

Changes

New (no impact on existing callers)

• agent/cli.py — memory subparser; cmd_memory_list/show/search/forget;
  _MEMORY_TYPES tuple; _MEMORY_TYPE_STYLES (with module-load assert);
  rich.markup.escape on every user-supplied field; Confirm.ask
  wrapped against EOFError for non-interactive deletes
• agent/src/memory/persistent.py — MEMORY_TYPES constant; module
  logger; public list_entries() / find(name) / remove_entry(entry).
  Existing remove(title) is unchanged so remember_tool is unaffected
• agent/tests/test_cli_memory.py — 29 new tests (command logic,
  name resolution incl. true stem fallback, parser wiring, Rich markup
  escape, empty body rendering, EOF path, remove-failure path,
  bracketed-description regression)

One existing-code change: _scan_entries coercion

MemoryEntry annotates title, description, and memory_type as
str, but _scan_entries was passing the raw parse_frontmatter
output through untouched. The parser returns lists for [a, b]
syntax and bools for true/false, so any markdown file with e.g.
description: [red]inject[/red] leaked a list into the dataclass.
The previous CLI never called string ops on these fields, so the
mismatch was latent; the new CLI hits rich.markup.escape(...) and
crashes with TypeError.

Add a small _coerce_str helper and route the three frontmatter
fields through it so the dataclass contract is truthful at the storage
boundary. Existing consumers (find_relevant's tokenization,
remember_tool's rendering, remove(title)'s title comparison) all
silently improve on the edge case and behave identically on the happy
path. tests/test_remember_tool.py (16 tests) is unchanged and still
green.

Pre-existing limitations surfaced by the new CLI (out of scope)

Heavy E2E testing with Thai content surfaced two pre-existing bugs in
persistent.py that this PR intentionally does not fix:

• _tokenize accepts only CJK Unified Ideographs (一-鿿);
  Thai / Arabic / Hebrew / Cyrillic queries always return zero
  matches because the tokenizer returns an empty set on those scripts
• The slug regex in add() strips Thai (and other non-CJK scripts)
  to _, producing filenames like feedback______________.md for
  Thai titles

These are follow-ups of #95 and worth a separate issue/PR. The new
CLI happens to make them visible (memory show renders the Thai
title fine, but the on-disk slug is mangled).

Test Plan

• pytest agent/tests/test_cli_memory.py agent/tests/test_persistent_memory.py agent/tests/test_remember_tool.py agent/tests/test_cli_init.py → 81 pass
• ruff check agent/cli.py agent/src/memory/persistent.py agent/tests/test_cli_memory.py agent/tests/test_persistent_memory.py → no new errors (14 pre-existing E402/F541/F401 untouched)
• time vibe-trading --version → 1.0s
• File-level E2E on a populated ~/.vibe-trading/memory/: list (with and without --type filter), show by title / by full slug / by bare slug, search (Latin and CJK queries), forget with -y, with n decline, with closed stdin (EOF)
• End-to-end with a real LLM: vibe-trading run → agent invokes remember tool → new entry visible in memory list, retrievable via memory show, matchable by memory search, removable via memory forget -y

Checklist

• No changes to protected modules (src/agent/, src/session/, src/providers/)
• No hardcoded values
• Follows CONTRIBUTING.md
• Tests added
• No on-disk format changes; existing remember_tool workflows are byte-for-byte compatible

[warren618]

Thanks @Teerapat-Vatpitak — this is a clean and useful addition. The memory CLI makes the persistent-memory layer much easier to inspect and debug, and the extra frontmatter coercion fix is a nice catch. Merged!