fix(memory): add Step 3.5 (codebase verification) to task_planning SOP (closes #560)

[Kailigithub]

Summary

Add Step 3.5 (现状核实 / codebase verification) to task_planning.md so that when the autonomous agent generates candidate TODOs in planning mode, it first verifies the candidate features/modules are not already implemented in the codebase.

Closes #560.

Root cause

task_planning.md Step 3 only inventories prior reports and memory artifacts (ls autonomous_reports/ + read ../memory). It does not check the actual code repository, which means already-implemented features get treated as "candidate new TODOs" of class 评估/集成 (evaluate/integrate). The agent then spends ~30 turns producing plan outputs that are not actionable.

Fix

Insert Step 3.5 between current Step 3 and Step 4. The new step instructs the agent to verify candidates with concrete shell commands (find . -name / grep -rln) before writing 评估/集成-class TODOs, and to forbid those classes entirely for already-implemented features.

Why this re-roll

The previous attempt (#561) was closed after 46 days without maintainer review activity, per the bot's housekeeping protocol. The original diff referenced a non-existent helper verify_custom_features; this v2 replaces that with concrete shell commands (find / grep) so the verification step is executable without any project-specific helper.

Diff size

+3 lines, single file, single commit. No external dependency, no behavior change for any runtime path — this is a documentation/SOP change only.

Verification

python3 /tmp/test_issue_560.py
# ✓ all checks passed: Step 3.5 (现状核实) is present, well-placed,
#   uses concrete shell commands, and constrains output-class TODOs

The test was confirmed to FAIL on the unmodified task_planning.md (Step 3.5 absent) and PASS with the fix applied (three-step dance).