Run LoCoMo end-to-end; establish baseline numbers

[StigNorland]

Once the dataset (#9) and spaCy (#8) are in place, run the full pipeline and record baselines. No results/ directory currently exists, so all thresholds are uncalibrated guesses.

• Run locomo eval across the 5 QA categories
• Create results/ and commit a baseline snapshot
• Compare graph vs. graph_only vs. baseline context modes

Blocked by #8, #9.

[StigNorland]

First runnable results (sample conv-26, 5 QA, via new src/run_locomo.py):

category	graph	baseline
1 retrieval	0.25	0.50
2 temporal	0.0	0.0
3 inference	0.75	0.75
overall	0.25	0.35

• Graph mode loses to baseline in every config; temporal (cat2) is 0.0 for both.
• Concept extraction is off by default in this path (use_turn_level_concepts=False); even with it on, graph<baseline (root cause → Calibrate dynamic maturity/merge thresholds (design gap #2) #16).
• Harness landed; this is a 5-QA smoke, not the full baseline. Full 10-sample run still TODO here.

[StigNorland]

Full baseline complete — sample 0 (conv-26), all 199 QA, concepts ON, via src/run_locomo.py. This is the pre-fix state (collapsed 3-hub concept graph), now the documented before for the merge fix (#16).

cat	graph	baseline
1 retrieval	0.117	0.086
2 temporal	0.041	0.000
3 inference	0.365	0.404
4 multi-hop	0.132	0.050
5 adversarial	0.830	0.968
overall	0.293	0.286

Key reads:

• At full scale graph mode edges ahead overall (opposite of the 5-QA smoke) and wins retrieval + multi-hop (2.6×).
• Loses inference + adversarial — the collapsed graph feeds noise, so it over-answers unanswerable/adversarial items it should decline.
• Temporal ~0 for both modes → bi-temporal + as-of is the highest-value v2 workstream.

Artifact: results/baseline_s0_full.json. Post-fix graph-only run (688 concepts, #16 fix) in progress.