PRIME C-19 - Phase-Recurring Infinite Manifold Engine (Project David)

Changed — 2026-04-25: branch consolidation + tools cleanup pass Repo-wide cleanup that merges the parallel research tracks back into a single main and prunes 53 legacy scripts from the previous research arc. Per the ARCHIVE.md protocol, every removed surface is preserved at a uniquely-named tag. Branch consolidation: 6 branches collapsed into single main. The codex/phase-b-logging-smoke Phase A→B sweep tooling (11 commits) was merged via 3-way merge with theirs/codex resolution on 4 add/add or content conflicts in evolve_mutual_inhibition.rs, diag_dimensionality_sweep.py, diag_phase_d0_5_jackpot_aperture.py, and docs/PHASE_B_PRE_REG.md (citation-precision fix on the last). The claude/review-repo-access-Ug8Si branch was verified fully redundant (codeql.yml blob hash identical, 7 LCF docs commits already in PR #146). The research/overnight-sct-empirical-20260423 8-iteration LOOP-COMPLETE branch and the two saved/* historical branches were archive-tagged and removed. Archive tags created: archive/main-pre-cleanup-20260425, archive/codex-phase-b-logging-smoke-20260425, archive/research-overnight-sct-empirical-20260425, archive/saved-neuron-one-wip-20260425, archive/saved-pre-connectome-research-20260425, archive/claude-review-repo-access-20260425 — six branch heads preserved as immutable snapshots. Tools/ trim: 74 entries → 22. Removed 52 scripts from the 2026-04-17 → 2026-04-19 byte-pair merger / Block C quantization / word-tokenizer research lines (Block C activation/quantization 14, byte-pair merger 10, byte Huffman/L2 8, word tokenizer 6, MLP baselines 4, probe-weight 2, modal/sweep 2, pretokenize 2, misc 4). Preserved at tag archives/tools-cleanup-20260425. Note: diag_byte_unit_widen_sweep.py was initially staged for archival but restored after a post-archive verification caught that build_byte_unit.py (a kept reproduction script) imports from it; preserved in tools/ as an internal dependency only. Net diff: −13,884 lines. tools/README.md rewrite: previous version listed canonical scripts (diag_qat_ste, diag_byte_pair_merger_lookup_codebook, diag_byte_single_w_* series) that had been archived in the 2026-04-18/20 cleanup passes; the file was stale relative to the actual tree. Rewritten to reflect the current 21-entry layout: public-beta contract scripts, CI utilities, Block A deploy artifacts, and the active Phase A/B/D research line. SDK comment updates: Python/block_c_embedder/embedder.py docstring + README.md, Python/block_b_merger/merger.py, and Rust/src/block_b_merger/mod.rs comments updated to reference the archive tag instead of dangling tools/ .py paths. Runtime is unchanged; only documentation strings move. .gitignore fix: instnct-core/target/ added to the rust-build-artifacts gate (the workspace-member target/ directory was previously untracked-but-not-ignored). Wiki sync: 5 LCF sub-pages created on the GitHub wiki (Local-Constructability-Framework.md, Interference-Dynamics.md, Mutation-Selection-Dynamics.md, Constructed-Computation.md, Cognitive-Emergence-Speculative.md); 4 existing pages updated (Home.md, Theory-of-Thought.md and Structured-Chaos-Theory.md as redirect stubs, _Sidebar.md for nav). Added — 2026-04-23/25: Phase A → B → D mutation-selection research line Multi-seed dimensionality / mutation-selection / acceptance-aperture study built on evolve_mutual_inhibition and evolve_bytepair_proj fixtures, replacing the single-seed observations from the late beta.2 era. Phase A baseline (docs/research/PHASE_A_BASELINE.md): H ∈ {128, 192, 256, 384} × 5-seed sweep on the byte-pair prediction fixture. Measured peak_acc, accept-rate, alive_frac, edges; revealed an inverted-U with peak at H=256 (mean 5.28% ± 1.79pp) and a monotonic accept-rate collapse 78→42→13% as H grows. Phase B confound-vs-intrinsic test (docs/research/PHASE_B_VERDICT.md, docs/PHASE_B_PRE_REG.md): pre-registered 5-arm × 5-seed protocol at fixed H=384 testing whether the H=384 decline is intrinsic or driven by step budget / jackpot size / propagation depth / input-channel saturation. Drives the C_K constructability metric on per-candidate logs. Phase B.1 horizon × accept-ties follow-up (docs/research/PHASE_B1_PRE_REG.md, docs/research/PHASE_B1_VERDICT.md): 3-tier horizon scan (S20 / S40 / S80) crossed with strict / ties acceptance; identifies the binding constraint as the acceptance valve at H=384. Phase D0/D0.5/D1 acceptance-aperture series: D0 acceptance-aperture metric on candidate logs (docs/research/PHASE_D0_ACCEPTANCE_APERTURE.md); D0.5 offline K-resampling (docs/research/PHASE_D0_5_JACKPOT_APERTURE.md) separates the jackpot/sampling aperture from the acceptance valve; D1 zero-drive policy K × zero_p factorial sweep. Driver consolidation in tools/diag_dimensionality_sweep.py: single multi-mode driver covering default H-sweep + --phase-b + --phase-b1 + --phase-d1 arms, with ThreadPoolExecutor-based per-cell parallelism (--jobs N). Phase B CLI on evolve_mutual_inhibition.rs: 14 new flags (--jackpot, --ticks, --accept-ties, --accept-policy, --neutral-p, --accept-epsilon, --input-scatter, --candidate-log, --checkpoint-at-end, --panel-interval, --panel-log, --phase, --arm, --run-id); AcceptancePolicy enum (Strict / Ties / ZeroP / Epsilon) plumbed through instnct-core/src/evolution.rs (+543 lines); RunMeta JSON serialization for reproducible artifact provenance; extended SUMMARY JSON with phase / arm / run_id / horizon_steps / accept_ties / accept_policy / neutral_p / accept_epsilon fields. Five new analyzers: analyze_phase_a_baseline.py, analyze_phase_b_verdict.py, analyze_phase_b1_verdict.py, analyze_phase_d1_verdict.py, analyze_acceptance_aperture.py; one new diagnostic diag_constructability_analysis.py (C_K decomposition regression across V_raw / M_pos / A / I_proxy / D_eff / cost_eval / R_neg). CodeQL workflow (.github/workflows/codeql.yml): pinned to {python, rust, actions} to drop the failing default c-cpp matrix job that was producing constant red CI. Added — 2026-04-21/22: ABC-Brain integration, fitness sweep, crystallize, ablation First end-to-end wiring of the frozen ABC feature pipeline into the INSTNCT brain. Intensive experimentation covering fitness function optimization, structural experiments, crystallize port, and ablation study revealing single-attractor topology collapse. ABC-to-Brain char-level integration: C-embedding matches SdrTable at ~25% (multi-seed validated). First end-to-end wiring confirms embedding quality is not the bottleneck — brain topology is. Byte-pair prediction (397 class): 7.1% peak with smooth cosine fitness (frequency baseline 4.2%). Real signal on a much harder task, but brain topology limits further progress. Fitness function sweep (10 variants): smooth linear cosine champion. Dominates stepwise, argmax, pure-accuracy, and other cosine variants. The fitness signal shape remains the single biggest lever for mutation-selection. Crystallize ported from Python to Rust: grow-prune-regrow cycles validated. Converges to compact circuits that retain functional accuracy. Ablation study: systematic ablation reveals single-attractor topology collapse — 7 dominant neurons form a bottleneck. The brain converges to one attractor basin instead of developing competing pathways. This is the core pathology limiting current accuracy. Structured Chaos Theory v1.0: three laws (Single Constraint, Anti-Monopoly, Opponent) formulated from accumulated experimental evidence. Learning formula: S x sensitivity / dimensions. Added to docs/wiki/Theory-of-Thought.md. Added (negative results) — 2026-04-21/22 Edge weights [1-3]: worse than binary — weighted edges degrade performance vs binary {0,1} masks. Signal-to-noise ratio degrades when edge precision increases. Reconfirms topology > edge precision. Multi-channel input: worse than single — dual-input and multi-channel injection schemes all worse. Dimension curse: extra channels increase search space faster than they add useful signal. Added — 2026-04-21: Block C byte-pair embedder champion + deploy SDK Canonical ABC-pipeline-ready Block C embedder trained, quantized, and packed. Full bytes-in / embeddings-out path now available as pure numpy. Training (Modal L4 GPU, ~$9 total): full-softmax next-pair CE on 100 MB FineWeb-EDU, E=32 / H=128 / context=16, 3 seeds (1, 3, 7). Two-phase: v1: LR=0.1 from scratch — diverged after ep3 (peak acc@1 31%), but revealed strong syntactic clusters. v2 (champion): warm-start emb from v1 ep3 + LR=0.03 + cosine decay, 10 epochs — acc@1 34.06 ± 0.82%, no divergence, clusters tightened and extended. Intelligent quantization (tools/diag_bytepair_mixed_quant.py): hot-vs-cold split by corpus frequency. 3,386 pairs w/ freq ≥ 5 quantized to per-channel int4 α=0.5; 62,150 cold pairs collapse to one shared OOV vector. Cluster overlap with float reference: 74.4% on hot-restricted top-5. Baked artifact: output/block_c_bytepair_champion/packed.bin (62,528 B / 61 KB, 134× compression vs float32 8.39 MB). Format: VCBP v1, per-channel scales fp16, shared OOV fp16, 65,536-bit hot bitmap, int4-packed hot rows. Bake script: tools/bake_block_c_bytepair.py. Python deploy SDK: Python/block_c_embedder/embedder.py — L2Embedder.load_default(), embed_id, embed_ids, encode_bytes. Zero ML deps. Chain A+B+C stress test (10 invariants, all pass): A round-trip lossless, B sign-match lossless, C header + scheme, C determinism, OOV sharing across 100+ cold pairs, hot uniqueness > 99.5%, semantic cluster preservation (. → ! /? /.\n, , → ; /: , ' t' → '\nt'/'(t'/'-t'), 100 KB corpus encode, edge cases (empty/single/odd/binary). Script: Python/block_c_embedder/tests/test_chain_a_b_c.py. Learned clusters (emergent from data, no hand-crafted labels): word-start equivalence (' t' = '\nt' = '(t' = '-t'), case-invariance (th↔Th, he↔He, in↔In, on↔On, an↔An), sentence-terminator group, clause-punct group, function-word group (in/by/on). Infrastructure: Modal app wrapper tools/modal_block_c.py with L4/T4/A10G