feat: persist ConversationContext across sessions (closes #164)

[luceinaltis]

Summary

Closes #164.

ConversationContext was previously rebuilt from the session log on every query and lost on process exit. This PR adds a durable layer so returning users keep their topic stack, last intent, and activity timestamp across restarts — the "welcome back" experience flagged as 구현 예정 in docs/user-experience.md.

What changed

• qracer/conversation/context_store.py (new) — pure helpers:

  • save_context / load_context — JSON serialization, tolerant of missing files, malformed JSON, non-dict payloads, bad ISO timestamps, and non-string topic entries.
  • decay_stale(ctx, max_age_days=7) — clears topic_stack/current_topic once the gap since last_activity exceeds the threshold. Because the data model carries a single last_activity shared by all topics, "decay stale topics" is implemented as "the whole stack is abandoned once the threshold is crossed" — documented in the module docstring.
  • merge_with_theses(ctx, thesis_tickers) — folds open FactStore theses into the stack (deduped, bounded by MAX_TOPIC_STACK), seeding current_topic from the stack if unset.
  • merge_contexts(session, persisted) — combines an in-session context with the persisted one: session wins on every "right now" scalar (current_topic, intent, depth, last_activity); the persisted topic_stack supplies trailing entries.

• qracer/conversation/engine.py — new context_path: Path | None = None kwarg. On init the persisted context is loaded, decayed, and enriched with FactStore.get_open_theses(). After each successful query the engine merges extract_context(turns) with the persisted context and writes the merged state back to disk. Load and save are both try/except-guarded so a bad file can never break the query loop.

• qracer/cli.py — repl wires context_path=_user_dir() / "context.json" so real users get persistence for free.

Scope mapping (issue #164)

Scope item	Status
Serialize on session end / compaction	✅ Saved after every query (covers both)
Store in a dedicated context.json in ~/.qracer/	✅ context.json
Load on session start, merge with FactStore open theses	✅ _load_initial_context
Decay stale topics (>7 days → remove from topic_stack)	✅ decay_stale, DEFAULT_STALE_DAYS=7

Test plan

• uv run pytest tests/conversation/test_context_store.py tests/conversation/test_engine.py — 68 passed (23 new test_context_store cases + 5 new TestPersistentContext engine cases).
• uv run pytest full suite — 806 passed, 14 skipped.
• uv run ruff check on changed files — clean.
• uv run pyright qracer/conversation/context_store.py qracer/conversation/engine.py — 0 errors.

New tests cover

• Save/load round-trip preserves all fields (current_topic, stack, intent, depth, last_activity).
• Missing file, malformed JSON, non-dict payload, bad timestamp, non-string topic entries all yield a safe default context.
• Fresh context preserved by decay_stale; stale context clears stack but keeps last_activity; custom max_age_days; empty-context fast-path.
• merge_with_theses appends, dedupes, respects MAX_TOPIC_STACK, seeds current_topic, preserves existing one.
• merge_contexts session-wins semantics, persisted fallback for missing scalars, topic-stack merge order (session first), MAX_TOPIC_STACK bound, last_activity carries session's value.
• Engine integration: context file written after a query, topics resumed in a fresh engine with the same context_path, stale context decayed on load, FactStore open theses merged into the stack, engine without context_path is a silent no-op.

Manual verification

from pathlib import Path
from qracer.conversation.context import ConversationContext
from qracer.conversation.context_store import save_context, load_context

ctx = ConversationContext(current_topic="AAPL", topic_stack=["AAPL", "NVDA"], intent="research")
save_context(ctx, Path("/tmp/ctx.json"))
load_context(Path("/tmp/ctx.json"))
# ConversationContext(current_topic='AAPL', topic_stack=['AAPL', 'NVDA'], intent='research', ...)