Build the v50 evidence-first research kernel

[DevHabito]

What this changes

• introduces the packaged darwin-v50 evidence-first research kernel
• separates the maintained v50 architecture from historical v47-v49 material
• documents the architecture, research protocol, legacy boundary, and contribution workflow in English
• adds packaging, CI, maintained-surface checks, and reproducible evaluation commands
• removes local runtime databases and exports from version control
• implements the H50-L1 through H50-L12 learning sequence with explicit causal archives and replay-checked snapshots

H50-L12 final decision

H50-L12 implemented online posterior-sampling control with exact tabular Beta posteriors, prequential order evidence, finite-horizon planning, paired potential outcomes, and equal interaction budgets.

The registered development seeds selected a 32-action resampling block. Final seeds 22100–22199 were run once. The hypothesis is refuted.

Three registered criteria failed:

• improvement over certainty-equivalent: -0.016421875, required >= 0.010
• improvement over epsilon-greedy: 0.0033671875, required >= 0.005
• simultaneous learned-baseline win rate: 0.04, required >= 0.60

Twelve other checks passed, including oracle ratios, model recovery, transition and reward error, archive retention, snapshot replay, and causal-field completeness. The final seed family is retired and will not be reused to promote a modified hypothesis.

Validation

• full suite: 219 tests collected; 218 passed, 1 skipped, 0 failed
• the skip is the Windows symlink path unavailable without the required OS privilege
• maintained-surface checker passed
• Python compile check passed
• all 11 declared console entry points resolve
• staged diff whitespace check passed
• H50-L12 final aggregate: docs/v50/results/EXPERIMENT_016_FINAL_AGGREGATE.json

Evidence boundary

This is an engineering and research foundation, not evidence of consciousness, sentience, subjective experience, AGI, or equivalence to Diana from Pragmata.

Current evidence remains at the project's E1 local automated evaluator level. H50-L11 and H50-L12 are both recorded as refuted; failed criteria remain visible.