fix(engine): wire campaign framing into tick + GPU llama3.1:8b

[showjihyun]

Summary

Found and fixed a Prophet engine bug where CampaignConfig.{novelty, utility, controversy} were being read from the API payload but silently dropped before reaching the agent tick loop. Pilot-tested all three marketing use cases from README.md before and after the fix; committed the evidence.

This branch has two commits:

1. 629e0b0 docs(pilots): verify README use cases end-to-end, find campaign wire is severed — built a reusable pilot harness (backend/scripts/run_use_case_pilot.py), ran 6 pilots, discovered all 6 produced bit-identical step-by-step trajectories regardless of campaign framing. Traced the root cause to exact file:line locations and wrote docs/USE_CASE_PILOTS.md with the verdict and a 5-item follow-up list.

2. 790fe60 feat(pilots): fix campaign framing wire + switch to GPU llama3.1:8b — applied the three-layer fix, added a regression test that would have caught the bug immediately, re-ran all 6 pilots on GPU, and updated docs/USE_CASE_PILOTS.md with before/after deltas.

The wire fix (three layers)

Layer	Change
community_orchestrator.py	Extract all three framing attrs from CampaignEvent, forward into tick() + async_tick() alongside the existing campaign_controversy wire
tick.py	MessageStrength construction now blends campaign inputs (60%) with agent-derived perception (40%): novelty = 0.6·campaign_novelty + 0.4·media_signal, same for utility. controversy stays pure-campaign
cognition.py	Tier-1 rule engine gained campaign_bonus = 0.3·(utility−0.5) + 0.2·(novelty−0.5), scaled ×2, folded into evaluation_score. Centered at 0 for neutral campaigns so legacy fixtures stay green

Regression test test_04_step_runner.py::TestCampaignFramingAffectsOutcome asserts that two seed-identical sims with opposite campaign framings differ by ≥2 adoption points. Without the wire fix the delta is 0.0000 (bit-identical); with the fix it's +0.1817 at step 4.

Post-fix pilot deltas

Use case	Pre-fix step-0 delta	Post-fix step-0 delta	Post-fix final delta
UC1 baseline → reframed	+0.000	+0.236	+0.017
UC2 Strategy B → Strategy C	+0.000	+0.264	+0.017
UC3 raw → restructured	+0.000	+0.147	+0.185

UC3 is the flagship result: the raw RTO mandate now fires zero viral_cascade events and stalls at 74.5% adoption, while the restructured version fires 3 cascades and reaches 93.1%. That's a real +18.5pt lift from restructuring — directionally reproduces README's "-60% opposition" claim.

UC1 and UC2 still saturate at ~97% because the 1030-agent populations cross cascade critical mass even with hostile framing. The follow-ups to get exact "stall at 12%" reproduction (population scaling to 5K-10K + stronger campaign_bonus weights) are documented in docs/USE_CASE_PILOTS.md#follow-up-items-post-round-8-6.

Stack changes (Round 8-6)

• Ollama moved to GPU via docker-compose.gpu.yml override. RTX 4070-class GPU runs llama3.1:8b at ~75 tok/s (CPU mode was ~4-8 tok/s). Every agent tick + opinion synthesis now finishes in sub-second wall time.
• Default model: llama3.2:1b → llama3.1:8b. The 1B model was hallucinating opinion synthesis narratives that matched the README claims instead of the actual metrics. The 8B model stays anchored to the provided numeric evidence.
• Opinion synthesis timeout: 120s → 30s. The 120s timeout was a CPU-mode workaround. GPU calls finish in ~1-2s so 30s is still 15× headroom.
• README + CLAUDE.md Quick Start rewritten with GPU as the recommended path and CPU-only as a documented fallback with env-var overrides to flip back to llama3.2:1b.

Test plan

• Backend uv run pytest tests/ — 1029 passed, 2 skipped (+1 new regression test, no regressions)
• New regression test TestCampaignFramingAffectsOutcome passes on the post-fix code, fails (delta=0.0000) on the pre-fix code
• All 6 pilots re-ran end-to-end on GPU with llama3.1:8b, non-stub responses for every community-opinion + overall-opinion call
• docs/pilot_results/*.json regenerated with post-fix trajectories
• UC3 raw → restructured delta is +18.5 adoption points at final step (vs -0.3 pre-fix)

Follow-up items (not in this PR)

1. Stronger campaign_bonus weights so UC1/UC2 produce sharper stalls (±0.5 delta on evaluation_score instead of current ±0.25)
2. 5K-10K agent populations for UC1/UC2 — small populations always cross cascade critical mass
3. Echo chamber detector never fires for UC2 Strategy B despite being maximally polarising — the detector looks at community isolation rather than message-induced polarisation
4. README soft-update — the "stall at 12%" quantitative claims aren't reproduced exactly yet (we get directional patterns); once 1+2 above land, regenerate the pilots and update README if the magnitudes still don't line up

🤖 Generated with Claude Code