feat: MEMORY.md / BOOTSTRAP.md cross-session long-term memory (closes #160)

[luceinaltis]

Summary

Closes #160.

docs/memory-system.md line 38 flagged MEMORY.md / BOOTSTRAP.md as 구현 예정 — Tier 2 summaries are auto-generated per-session but there was no durable bridge between them and the user-curated long-term context. This PR adds that layer.

• MEMORY.md at ~/.qracer/MEMORY.md — a single Markdown file with a machine-managed auto region delimited by <!-- BEGIN:auto --> / <!-- END:auto -->. The auto region holds open theses (with entry/target/stop/catalyst) and upcoming catalysts, regenerated from FactStore after every thesis save. Everything outside the auto region is user-curated free text and preserved verbatim across refreshes.
• BOOTSTRAP.md at ~/.qracer/BOOTSTRAP.md — optional raw text loaded once at ConversationEngine init and pushed onto the session history as a system turn so long-term preferences ("I'm a long-term value investor") reach the synthesizer without code changes.

What changed

• qracer/memory/memory_file.py (new) — self-contained module:

  • MemoryDocument dataclass (auto_theses, auto_catalysts, user_content, last_updated) plus summary_line() for briefings.
  • render_memory(doc) / parse_memory(text) round-trip.
  • load_memory(path) / save_memory(doc, path) with atomic .tmp-swap writes and tolerant fallbacks for missing/unreadable/malformed files.
  • refresh_memory(doc, fact_store, catalyst_horizon_days=30) — pure function that regenerates the auto region from FactStore.get_open_theses() and get_upcoming_catalysts() without mutating user content.
  • refresh_memory_file(path, fact_store) — convenience wrapper.
  • load_bootstrap(path) — returns None for missing/empty files.

• qracer/conversation/engine.py — two new optional kwargs (memory_path, bootstrap_path). _prime_long_term_memory() runs at the end of __init__: BOOTSTRAP.md content becomes a system turn; MEMORY.md with non-empty auto content becomes a rendered system turn. _persist_facts now calls _refresh_memory_file() after a successful thesis save so the auto region stays current. All load/save paths are try/except-guarded — a malformed MEMORY.md can never break the query loop. New memory_document property exposes the cached doc.

• qracer/cli.py — repl() wires memory_path = _user_dir() / "MEMORY.md" and bootstrap_path = _user_dir() / "BOOTSTRAP.md" into the engine, and surfaces ✓ MEMORY.md: N active theses, M upcoming catalysts on startup when a doc loads. _repl_loop accepts memory_path and handles three new commands:

  • memory show — print the file (or a hint if absent).
  • memory refresh — refresh_memory_file() and sync the engine's cached doc.
  • memory edit — launch $EDITOR (falls back to nano/vi), seeding the file with a canonical template on first use.

  BANNER + _HELP_TEXT updated accordingly.

• docs/memory-system.md — replaces the 구현 예정 banner with the new format spec and CLI command list.

Scope mapping (issue #160)

Scope item	Status
Define MEMORY.md format (structured Markdown for theses, watchpoints, user prefs)	✅ auto/user split via BEGIN/END markers; canonical template on first memory edit
Load MEMORY.md at session start and inject into ConversationContext	✅ injected as a system turn in _prime_long_term_memory
Auto-update MEMORY.md when FactStore theses change (open/close/supersede)	✅ _refresh_memory_file runs after every _persist_facts save — supersession flips happen inside FactStore.save_thesis, so the regenerated open-theses list reflects them on the next render
BOOTSTRAP.md as a lightweight system prompt extension loaded before first query	✅ load_bootstrap + system-turn injection
CLI commands: memory show, memory edit, memory refresh	✅ all three wired

Test plan

• uv run pytest tests/memory/test_memory_file.py tests/conversation/test_engine.py — 81 passed (24 new test_memory_file cases + 7 new TestLongTermMemory engine cases).
• uv run pytest full suite — 809 passed, 14 skipped.
• uv run ruff check on changed files — clean.
• uv run pyright qracer/memory/memory_file.py qracer/conversation/engine.py — 0 errors, 0 warnings.

New tests cover

memory_file

• Render: empty doc renders placeholders; theses/catalysts become bullet lines; user content lives after the auto region.
• Parse: round-trip preserves every field; empty input → default; missing auto region preserves user body; malformed *Last updated:* falls back to now(); bullets outside the ## Active Theses heading are ignored; user content between auto region and trailing sections is captured.
• Refresh: auto region regenerated from FactStore; user content preserved; upcoming catalysts within horizon included; catalysts outside horizon excluded; last_updated bumped.
• Persistence: save/load round-trip; save_memory creates parent dirs; missing file returns default; refresh_memory_file preserves user section on disk; .tmp file cleaned up after atomic swap.
• BOOTSTRAP: missing returns None; empty file returns None; non-empty returns stripped content.
• summary_line counts.

Engine integration

• BOOTSTRAP.md content seeds a system message in history.
• Missing BOOTSTRAP.md leaves history untouched.
• Pre-existing MEMORY.md is parsed into memory_document.
• MEMORY.md with theses becomes a system message in history.
• Empty MEMORY.md is not injected.
• After a thesis query, MEMORY.md on disk is refreshed and engine.memory_document reflects the new thesis.
• Malformed MEMORY.md does not crash engine construction.

Manual verification

from pathlib import Path
from qracer.memory.fact_store import FactStore
from qracer.memory.memory_file import (
    MemoryDocument, refresh_memory_file, load_memory, save_memory
)
from qracer.models.base import TradeThesis

store = FactStore()
store.save_thesis(
    TradeThesis(
        ticker="AAPL", entry_zone=(175, 180), target_price=200, stop_loss=165,
        risk_reward_ratio=2.5, catalyst="AI revenue", catalyst_date="Q2 2026",
        conviction=8, summary="Long AAPL on AI.",
    ),
    session_id="demo",
)

path = Path("/tmp/MEMORY.md")
save_memory(MemoryDocument(user_content="## Mine\n\nKeep me."), path)
refresh_memory_file(path, store)
print(path.read_text())
# Auto region now lists AAPL; "Keep me." still present in user section.