Roundtable Pai v1.3.1 test report

Scope
- Restored original A/B/C/D/E product semantics
- Kept OpenClaw natural-language single-entry controller
- Verified one-turn-at-a-time progression and state persistence

Semantics restored
- A: agree with participant A in the current round
- B: agree with participant B in the current round
- C: agree with participant C in the current round
- D: stay silent and let discussion continue
- E: user has something else to say

Key fixes
1. Restored user-facing option block in references/user-facing-output.md
2. Restored SKILL.md constraints so the model must keep original A/B/C/D/E meaning
3. Updated controller choice parsing and state transitions:
   - A/B/C -> agree with slot A/B/C and advance exactly one round
   - D -> silent continue and advance exactly one round
   - E -> enter awaiting_free_speech and wait for user content
   - natural full sentence at choice stage -> treated as E/content directly
4. Removed v1.3.0-added deep_question / replace_request option semantics from active flow
5. Updated README and interface contracts to match restored semantics
6. Kept fixed rarity weights and explicit runtime state persistence

Automated tests run
- python3 -m unittest tests.test_controller
- Result: 6 tests passed

Covered paths
- start question -> candidate pool
- pick 3 participants -> first round directly
- E -> wait for user speech -> next round only once
- D -> silent continue -> next round only once
- natural sentence at choice stage -> treated as E/content
- reaching max rounds -> final conclusion, then new question starts next discussion
