fix(cli): read and write the decision log on the CLI path

[newnewself]

Problem

The persistent decision log is only touched inside propagate() → _run_graph(). The CLI (analyze → run_analysis) builds its own initial state and streams graph.graph directly, so on a normal tradingagents run:

• no pending entry is written;
• prior entries are never settled, so no realised return / alpha / reflection is produced;
• the Portfolio Manager prompt never receives past_context.

README: "The decision log is always on. Each completed run appends its decision to ~/.tradingagents/memory/trading_memory.md. On the next run for the same ticker, TradingAgents fetches the realised return … and injects the most recent same-ticker decisions plus recent cross-ticker lessons into the Portfolio Manager prompt." None of that happens on the primary entry point today.

This is the same shape as the --checkpoint bug fixed in #1249: lifecycle logic that lived only in propagate() was a no-op on the CLI. That fix introduced begin_checkpoint / checkpoint_input / end_checkpoint so both paths share the lifecycle; this PR does the same for the memory log.

Fix

• TradingAgentsGraph.prepare_memory_context(ticker, date) — settles pending entries for the ticker and returns the date-gated past-context block (Decision-log lessons lack a point-in-time guard — possible lookahead in backtests #1251 semantics preserved).
• TradingAgentsGraph.record_decision(ticker, date, final_state) — appends the decision; a run that produced no final decision (interrupted stream) is skipped with a warning instead of raising.
• propagate() now uses both; run_analysis() calls prepare_memory_context before building state and record_decision after a clean stream. A mid-stream failure records nothing, so the checkpoint is kept for resume and no partial decision is logged.

Tests

New tests/test_cli_decision_log.py:

• the shared methods append a pending entry, skip an empty final state, and carry resolved lessons into the context block;
• a run_analysis run with a fake graph and patched UI asserts the initial state receives past_context and the completed run is recorded.

Full suite: 658 passed, 1 skipped.

[Yijia-Xiao]

Thanks @newnewself. Implemented in 4a9f196: the CLI and propagate() share one decision-log path.